The Sears catalog arrived sometime before Christmas and was approximately the size and weight of a municipal phone directory, back when municipalities apparently believed every citizen should be issued enough paper to stun a burglar. My two brothers and I would take it to the floor and immediately flip to the back, where Sears kept the toys.

Why the toys were at the back I don’t know. Perhaps the company understood children and realized that no eight-year-old was going to patiently work his way through Ladies’ Foundations, Automotive Accessories, and Men’s Permanent Press Slacks before arriving at the slot-car sets.

We approached the catalog from the rear.

Our instrument was a red pen.

This turned the Sears catalog into something between a Christmas list and a crime scene. We circled what we wanted. Sometimes we circled the circle. If the item was particularly important, you might go around it four or five times, pressing harder with each revolution until the paper became embossed and the toy appeared to have been identified by police as the victim.

There were three of us, so diplomacy was required.

You couldn’t simply circle everything. Well, you could, but then your parents would recognize that you had no principles.

You needed a platform.

I want the chemistry set.

Why?

Because it’s educational.

Suddenly you were Jonas Salk.

My parents would come over periodically and inspect our work. This was when you discovered that adults could look directly at the thing you wanted more than anything in the world and fail to understand it completely.

“You already have one of those.”

No.

You had something belonging to the same broad category.

This was like telling someone who wanted a dog that he already had a mammal.

There were trucks, trains, games, bicycles, models, chemistry sets, telescopes and things involving electricity that would probably now require a waiver and protective eyewear. Every page presented another possible future self. With the telescope you became an astronomer. With the chemistry set, a scientist. With the walkie-talkies, an intelligence operative communicating with another intelligence operative approximately thirty feet away.

We circled our futures in red.

I don’t remember ever wondering where these things actually came from.

That is one of the peculiarities of childhood. The world is constantly producing objects on your behalf, and you don’t feel much obligation to investigate the supply chain.

Oranges came from the grocery store. Gasoline came from the gas station. Christmas presents came from Christmas.

The Sears catalog was particularly mysterious because the thing you were looking at wasn’t there. It was a picture of a thing that existed somewhere else.

Where?

I suppose if you’d asked me I might have said Chicago. Sears sounded like something that came from Chicago.

Or perhaps there was a great national toy repository somewhere. I don’t think I would have ruled out the North Pole.

What I did not know was that a substantial portion of this mysterious elsewhere was sitting on Lawndale Drive.

I went there.

I went there several times.

I just didn’t know what it was.

The Sears Warehouse sat in a peculiar patch of Greensboro, an enormous windowless brick building associated with railroad tracks and loading docks, the sort of place you could live three blocks from for twenty years without ever wondering what happened inside.

You didn’t shop there, exactly.

For shopping, you went to the showroom, where merchandise was arranged in little domestic tableaux. A lamp stood beside a sofa. A television faced two chairs. A washing machine waited with its lid open as though expecting someone to put a sock in it.

My parents would examine whatever they were considering buying.

My father would open things.

This seemed to be his principal method of shopping.

Open the refrigerator, the washing machine, the cabinet, the dishwasher. If possible, crouch down and inspect something underneath.

My mother asked questions.

Eventually a salesman wrote something on a piece of paper.

Then, instead of taking the thing home, you got back in the car and drove around the building to the Sears Warehouse. You had to drive because the building was over 1.5 million square feet.

There you pulled up to a loading dock and handed another man another piece of paper.

The man disappeared into the building. Eventually he returned with your lamp.

It was an extraordinary system when you think about it. Sears had separated shopping from receiving the thing you’d shopped for and placed the two activities in different buildings.

Sometimes the man disappeared for a long time.

My parents experienced this as an inconvenience.

I experienced it as an opening.

I would get out of the car and go inside.

This was apparently permitted in the sense that nobody successfully prevented it.

I don’t remember my parents saying, “Go explore the enormous industrial warehouse by yourself.”

On the other hand, I don’t remember them saying, “Under no circumstances enter the enormous industrial warehouse by yourself.”

There was a great deal of useful territory between these two statements in the 1970s.

Inside, the building made no architectural sense to me.

The floors were too close together, maybe three-quarters the height of ordinary floors, with little stairways connecting them. Years later, when I encountered the compressed stacks in old research libraries, I recognized the sensation immediately: someone had decided that human beings were an inefficient use of vertical space.

The warehouse was designed for objects.

We were merely visiting.

There were very few windows, perhaps none in the places I wandered, and the lighting was dim enough to give everything the appearance of having been remembered rather than seen.

You would climb six steps and encounter dining room chairs.

Another flight: mattresses.

Turn left: lamps.

Another staircase: boxes.

Then washing machines.

Then rugs.

Then something you couldn’t identify at all because there were four hundred of them wrapped in brown paper.

That was the thrilling part.

A store contained things.

The Sears Warehouse contained quantities.

The showroom had a lamp.

The warehouse had lamp.

Lamp as a geological formation.

You would turn a corner and find yourself facing enough identical end tables to furnish a small dictatorship.

There were appliances in military formation. Mattresses rose in soft cliffs. Cardboard boxes made streets and alleys. Somewhere a forklift would beep, which I understood only as evidence that there were adults elsewhere in the ecosystem.

I was gloriously, intentionally lost.

Getting lost was the activity.

Today, getting lost has been almost completely ruined. Your phone knows where you are. Your parents know where you are. A satellite knows where you are. Somewhere a server farm can probably reconstruct the precise route you took through a Walgreens.

In the Sears Warehouse, nobody knew where I was.

Including me.

And I wasn’t frightened.

This seems strange to me now.

The building was enormous. I was a child. There were no signs saying YOU ARE HERE, and if there had been, I wouldn’t have known where “here” was in relation to anything else.

But I had a system.

Walk in one direction until you hit a wall.

Follow the wall.

Eventually something will happen.

A corner. A staircase. A door. Another wall.

This struck me as mathematically sufficient.

And it was.

I never remained in the Sears Warehouse permanently.

Though perhaps someone did.

Perhaps there is a boy who went in around 1974 and has been living ever since behind Major Appliances, emerging at night to eat packing peanuts and wondering why nobody has come to get him.

The oddest thing is that I don’t remember ever finding the toys.

You would think I would have.

I found everything else, or at least everything a child had no conceivable interest in finding. There were regions of the building devoted to objects whose existence had never previously occurred to me. You could stumble upon an entire population of lampshades.

Not lamps.

Lampshades.

Hundreds of them.

A lampshade is unremarkable in a living room. Put three hundred lampshades together in a dim warehouse and they begin to suggest a religious ceremony.

But never the toys.

At least I don’t remember them.

And here’s the part I only discovered recently.

That building wasn’t simply a Sears warehouse.

It was a Sears catalog distribution center.

The distinction would have meant nothing to me then.

It means everything to me now.

Because while my brothers and I were lying on the floor at home, red pen in hand, circling toys in the Sears catalog, some of those toys may already have been sitting on Lawndale Drive.

They weren’t in Chicago.

They weren’t in some great mysterious Christmas depot.

They weren’t at the North Pole.

They might have been two miles away.

And on some Saturday afternoon, while my parents waited at the loading dock for a lamp, I may have been wandering through the same building where my Christmas presents were waiting.

I might have passed within twenty feet of the bicycle I’d circled.

Maybe it was one floor above me.

Maybe the chemistry set was on the other side of a wall.

Maybe somewhere in that enormous building was a shelf containing the exact toy whose photograph my brothers and I had nearly carved out of the catalog with a red Bic.

I love that I didn’t know.

Knowing would have spoiled it.

If someone had said, “The toys are in Section 14-B,” I would have gone to Section 14-B.

Instead there was the possibility that they were everywhere.

Somewhere beyond the mattresses.

Behind the washing machines.

Past the lampshades.

Christmas had an actual physical location, and by complete accident I was allowed to wander around inside it.

Years later I saw Citizen Kane.

At the end of the movie, after Kane dies, his possessions are gathered in an enormous warehouse. Paintings, statues, furniture, crates, the accumulated material evidence of one fantastically acquisitive life, stretching away into darkness.

The camera makes Kane’s possessions look infinite.

I remember thinking:

Pssh.

You should see Sears.

Kane had one man’s lifetime.

The Sears Warehouse had everybody’s.

It had the washing machine your mother inspected and the television your father opened from the back for reasons known only to fathers. It had the lamp somebody would buy next Tuesday and the mattress some newlywed couple hadn’t yet decided they needed.

And somewhere, apparently, it had Christmas.

The building is what I think about now, more than the toys.

I couldn’t tell you what I circled most years. Whatever it was, I must have wanted it desperately. Then I got some of it, didn’t get the rest, played with what I got, outgrew it, broke it, lost it, gave it away.

The wanting was enormous.

The things turned out to be things.

But the warehouse remains gigantic in my memory.

Dark staircases. Low ceilings. Mountains of anonymous merchandise. The distant sound of a forklift. A little boy walking deliberately away from the loading dock until he no longer knew where he was.

Then walking farther.

Not because he was brave.

There was nothing to be brave about.

He knew exactly how the world worked.

Eventually you hit a wall.

Then you followed it.

And sooner or later, you found your way home.

Copyright © 2026 by Paul Henry Smith

You have been presented with a difficult constitutional question.

Do not panic.

The most important thing is to avoid understanding it too quickly.

STEP 1: READ WHAT CONGRESS WROTE

For example:

“The Postal Service shall receive, transmit, and deliver” the mail.

Ask yourself:

Do I understand what this means?

◻ NO

Congratulations.

You may already be Supreme Court material.

◻ YES

Careful.

You have fallen into the Plain Meaning Trap.

Proceed immediately to Step 2.

↓

STEP 2: IDENTIFY THE THING THE GOVERNMENT IS DOING

Suppose the government says:

If you do not satisfy our new requirements, we will not accept your ballots for mailing.

Ask:

Does “not accept” appear to conflict with “shall receive”?

◻ NO

Excellent. Skip to Step 7.

◻ YES

You are still thinking like a normal person.

Continue.

↓

STEP 3: LOOK FOR A DISTINCTION

Remember:

Two things that appear identical to an ordinary person may become completely different things if you are wearing a robe.

For example:

RECEIVING MAIL

and

ACCEPTING MAIL

At first these may seem suspiciously similar.

This is because you have not yet developed sufficient jurisprudence.

↓

STEP 4: ADD A PREPOSITIONAL PHRASE

This is one of the most powerful techniques in constitutional law.

The Postal Service is not refusing to receive mail.

It is:

establishing conditions for the acceptance of mail pursuant to its authority over the administration of the postal system.

Notice what has happened.

The original problem is gone.

Nobody knows where it went.

↓

STEP 5: CHECK WHETHER CONGRESS AUTHORIZED THE THING

Did Congress specifically say:

The Postal Service may refuse election ballots unless state election officials comply with a federally prescribed voter-information and barcode system?

◻ YES

Case closed.

◻ NO

Do not make the amateur mistake of concluding that Congress did not authorize it.

Instead ask:

Did Congress authorize something else that can be described broadly enough to include it?

For example:

“administer the postal system.”

There we are.

↓

STEP 6: REPHRASE THE QUESTION

This is crucial.

Never decide the question the irritating person is actually asking.

BAD QUESTION:

Can the Postal Service refuse ballots Congress told it to receive?

BETTER QUESTION:

May the Postal Service establish reasonable conditions governing entry of particular mailpieces into the federal mailstream pursuant to authority delegated by Congress?

Much better.

The answer has already begun to sound like yes.

↓

STEP 7: DISCOVER COMPLEXITY

Someone may object:

“But the statute says shall receive and the Postal Service says will not accept.”

Smile patiently.

Explain that the issue is more complicated than that.

Do not specify why immediately.

“More complicated than that” is the legal equivalent of putting your hand over someone’s mouth.

↓

STEP 8: ADD PRECEDENT

Find a case from 1894 concerning the mailing of lottery advertisements.

Add another concerning postage rates for agricultural newsletters.

Add a third involving a steamship.

The important thing is not that these cases answer the question.

The important thing is that there are now cases.

The obvious has become precedent-dependent.

We are making excellent progress.

↓

STEP 9: CREATE A TEST

A good legal test should contain at least three “prongs.”

For example, ask whether the government’s action:

1. falls within the historical ambit of postal administration;

2. bears a reasonable relationship to the integrity of the mailstream; and

3. does not unduly intrude upon a constitutionally committed state function.

You have now transformed:

CAN THE POST OFFICE REFUSE THE MAIL?

into:

WHAT IS THE HISTORICAL AMBIT OF THE MAILSTREAM?

This is why they gave you keys to the building.

↓

STEP 10: WEIGH THINGS

Courts love weighing things.

Put “state sovereignty” on one side. Put “postal integrity” on the other. Consider “administrative burden.” Add “public confidence.” Stir in “historical practice.” Mention “the separation of powers.”

Nobody knows what anything weighs.

That is fine.

You are not using a scale.

↓

STEP 11: RETURN TO THE ORIGINAL SENTENCE

Several hours later, someone will make the unfortunate observation:

“But it still says ‘shall receive.’”

This person is dangerous.

Explain that statutes must not be read “in isolation.”

Nothing in American law may ever be read in isolation, particularly if reading it in isolation causes it to become comprehensible.

↓

STEP 12: REACH YOUR DECISION

You are now ready.

You began with:

SHALL RECEIVE

and:

WILL NOT ACCEPT

These appeared to be opposites.

But after examining statutory context, historical practice, administrative authority, institutional competence, constitutional structure, relevant precedent and the historical ambit of the mailstream, you have discovered that the matter is considerably more nuanced.

Congratulations.

You have successfully thought like a Supreme Court justice.

ADVANCED EXERCISE

Try the following at home:

Congress:

“Do X.”

Government:

“We are not going to do X.”

Is this legal?

❌ BEGINNER: No.

❌ INTERMEDIATE: It depends.

✅ SUPREME COURT: The question improperly assumes that Congress’s use of “X” encompasses the particular species of X presently before us.

Welcome to the Court!

Copyright © 2026 by Paul Henry Smith

I have decided to write 1,000 symphonies.

This is the sort of announcement that makes people look not at you, but slightly to the left of you, as if there might be a responsible adult standing nearby who can take over the conversation.

“What do you mean by symphonies?” they ask.

By which they mean, “Please tell me you mean something else.”

I do not mean something else.

Over the next month or so, I am going to use generative AI to make an unreasonable amount of orchestral music. Then I am going to listen to it, reject nearly all of it, save the bits that seem alive, make more things from those, reject nearly all of those, and continue until either I find something worth keeping or I am found days later in my office, softly muttering “second bassoon” into a mug.

It is not a metaphor, though I expect to miss metaphors by Symphony 430.

The plan has the general shape of agriculture, if agriculture consisted of firing seeds into a ceiling fan and then crawling around with tweezers looking for wheat.

Still, I would like to know.

Can you make 1,000 symphonies and find one good one?

Or is this merely a new way of proving that volume is not the same thing as meaning, a lesson I could have learned by visiting Costco?

To make the experiment worse, or better, depending on whether you are a normal person or me, I am going to use the wrong AI model.

Not the latest Suno model.

That would be too sensible. It would be like showing up to a knife fight with a dentist.

I am going to use Suno 4.5, an older model that is about to be retired. Suno has announced that when the newer generation fully arrives, the old ones will go away. You will still be able to listen to things you already made, but you will not be able to make new things with them.

This gives the project the pleasant feeling of looting a hotel before demolition.

I am going to make 1,000 symphonies with a machine that is being led behind the barn.

Already this sounds more artistic than it deserves to.

The strange thing is that Suno 4.5 is not some primitive little steam-powered uncle of a model that I have grown attached to because it coughs charmingly before producing a waltz. When Suno introduced it, they emphasized its range. Genre combinations. Emotional range. The ability to respond to detailed musical prompts.

That is exactly what I want.

What exactly is Suno 4.5?

To be fair, and because every lunatic manifesto should include evidence, I could play people two Suno 4.5 songs just so they can hear how strange this thing is when left unsupervised near the silverware.

One is called “Chain Gang Stravinsky.”

The other is called “I’m in Love with the Chickens.”

These are not titles that emerge from a healthy institutional process.

They sound less like songs than things shouted by a man being removed from a conservatory.

And yet that is the point. I do not want the system that always arrives wearing a clean shirt. I want the one that sometimes shows up with feathers in its hair and a working knowledge of Petrushka.

I do not especially care whether the violins sound more like violins. I do not care whether the cymbals have a little electrical mildew around the edges. I do not care if the bassoons sound as though they were recorded in a municipal aquarium.

I am writing music.

If anything survives, musicians can play it.

This seems worth saying because generative music has inherited one of the stupidest confusions in the music industry: that sound quality and musical quality are somehow related. They are roommates at most. They share a refrigerator. They do not need to be married.

A magnificent performance from 1957 does not become less magnificent because the tape hisses. A boring performance from yesterday does not become profound because it was recorded at a resolution high enough to reveal the violist’s dental history.

The orchestral recordings I return to most often come from the 1950s, ’60s and ’70s. Reiner. Szell. Klemperer. Bernstein. Mravinsky. Solti. Karajan back when Karajan still sounded like a man who might make an unsafe decision with a sports car.

There is an astonishing variety in those recordings. Tempos that would now cause committees. Attacks that sound nearly criminal. Wind sections made of identifiable human beings instead of a single expensive beige sauce.

I am not saying civilization peaked and then misplaced the oboe.

I am saying that variety matters.

I want a model that can wander. I want it to understand that orchestral music does not mean “large tasteful movie cloud.” A phrase can be noble, vulgar, stiff, drunk, ceremonial, frightened, funny, provincial, overcooked, cruel, glassy, ridiculous, or French in a way no one can quite prosecute.

I want Mahler in one room, Haydn in another, and something suspicious behind the garage holding a bass clarinet.

The newer models may be better in every official way. Cleaner audio. More coherent structure. Better production. More control. More polish. They may be excellent according to the metrics that matter if your goal is to produce something that sounds like it could be released this afternoon and forgotten tomorrow morning.

That is not my goal.

I am not looking for commercially plausible orchestral music. We have plenty of that. It arrives every day in vans.

I want a machine capable of embarrassing itself.

I want the weird cousin.

I want the thing that gives me the articulation of a 1962 Czech orchestra, the proportions of a Classical finale, the emotional temperature of a nervous French ballet, and one brass gesture that makes everybody in the room glance toward the exits.

The audio can be dreadful.

I will survive.

Beethoven’s Fifth does not become a better composition because somebody remasters it.

So: 4.5.

The condemned model.

And then I begin.

Not by asking for 1,000 complete symphonies. That would be like adopting 1,000 horses and then asking whether anyone has seen my apartment keys.

I will start with openings.

Maybe a hundred.

A few seconds each.

No jury.

No solemn spreadsheet with columns labeled Emotional Integrity, Structural Novelty, and Bassoon Event Risk.

No rubric asking whether the piece “successfully communicates its intention,” as though a symphony were a memo from Human Resources.

I will listen like a man opening oysters.

No.

No.

No.

No.

Wait.

No.

Interesting.

God, no.

Maybe.

What the hell is that?

Keep.

That last category is important.

What the hell is that?

I suspect it may be more valuable than “excellent.” Excellence is easy to recognize once everyone has agreed what excellence sounds like. What I am looking for is the thing that briefly jams the machinery of expectation.

Those survive.

Maybe there are twelve.

Each of those gets children.

Twenty versions.

Now I have 240 things.

Cull.

Cull.

Cull.

This will be barbaric.

People have spent months working on eight bars of orchestral music. I will be throwing away eight bars every nine seconds while drinking coffee and wondering whether there is still yogurt.

Next.

Next.

Dead.

Keep the rhythm.

Lose the harmony.

Next.

Interesting middle.

Terrible beginning.

Save.

Next.

Someone call the authorities.

Next.

Then I start breeding the survivors.

An opening with a peculiar rhythmic infection gets extended. A harmonic turn from one thing gets transplanted into another. A movement that has become too smooth gets damaged on purpose. A beautiful melody that explains itself too easily gets sent back into the machine with instructions to become less employable.

Twenty continuations from the same thirty seconds.

Nineteen may be trash.

One may suddenly turn left.

Keep left.

Now make ten versions of left.

Generate.

Cull.

Branch.

Cull.

Mutate.

Cull.

Resurrect.

Regret resurrection.

Kill again.

At some point I expect to become emotionally attached to something obviously bad. This happens with children too, though apparently the rules are different.

Eventually the survivors become longer.

Thirty seconds.

Two minutes.

A movement.

And then the standard changes.

A twelve-second idea can be fascinating just by existing. At eight minutes, it has to have consequences. Did something come back? Did something change? Did the beginning ask a question the middle remembers? Is there any reason, other than sunk cost and mild dehydration, that I am still listening?

This is where generative music’s biggest danger may appear.

Not badness.

Agreeableness.

Perfectly lovely event.
Perfectly lovely event.
Perfectly lovely event.
Perfectly lovely event.

Seven minutes later I have apparently heard music, though there is no evidence.

A hotel lobby with excellent counterpoint.

And so, those die too!

And this brings me to the little bedtime story we are currently telling ourselves about AI.

But We have judgment and taste!

Yes, the machines can generate endlessly, we say, but humans will supply the judgment.

That is where we come in.

Taste!

Discernment!

Curatorship!

The machine generates. The human decides.

This is a comforting arrangement because it leaves humanity in a managerial position. We may not be digging the ditch anymore, but we are standing nearby in a good jacket saying, “Too wide.”

My 1,000 symphonies may demonstrate a small problem with this.

We won’t have time.

If I can generate a hundred openings over breakfast, what does “human judgment” mean?

If I can produce 10,000 candidate movements, who listens to them?

If a million songs arrive every morning, where exactly are all these sensitive curators supposed to stand? Is there a warehouse? Do they get stools?

Judgment is treated as though it were an unlimited natural resource, like air, or men on podcasts explaining optimization.

It is not. Judgment requires attention. Attention requires time.

Generative AI has removed the corresponding limitation on supply. And the arithmetic is not flattering.

It is not that humans will lose the ability to recognize something wonderful. It is that wonderful thing number 8,472 may get twelve seconds while we are looking for the AirPods case.

That is the experiment inside the experiment.

I am not John Henry racing the steam drill to see who can make music faster. That race is already over. The machine wins before I have found a chair.

I am John Henry trying to listen faster.

Send it.
No.
Another.
No.
Another.
No.
Another.
Wait.
Again.

The machine never gets tired.

I will.

The machine will not begin choosing stranger pieces after lunch because it is bored with C major.

I probably will.

The machine will not hear its seventieth opening and think, if one more French horn plays a rising fourth, I am going to throw this laptop into traffic.

I probably will.

My judgment will drift. My standards will change. I will mistake novelty for quality. I will miss things. I will save pieces in the morning that I would reject at night. I will become impatient. Somewhere around candidate 3,000, I may discover that the human capacity for aesthetic judgment is closely related to blood sugar.

And then, inevitably, someone will offer an AI system to fix this. It will listen to the 10,000 candidates for me. It will discard 9,950. Another model will rank the remaining fifty. Another will identify the outliers. And still another will learn my taste.

Perhaps it will notice that I keep saving passages with unstable middles, late harmonic turns, and bassoons behaving suspiciously.

Soon it will hand me six.

And, just like that, there goes judgment too!

That charming little preserve where humanity was supposed to sit in a cardigan lasted about as long as the free cheese at a gallery opening.

Which is why I want to do this stupidly. By hand. By ear.

While I still can.

One obsolete music generator on one side. One increasingly tired conductor on the other.

Go!

Eventually, perhaps, I will have twenty movements. Then twelve. Then five. Perhaps three plausible symphonies.

At that point musicians are needed.

Because musicians are a completely different evaluation system.

A passage that seems extraordinary through headphones may become instantly ridiculous when eighty people have to play it. Something I nearly deleted may acquire physical force the moment bows touch strings. Players may lean forward. Someone may ask to try it again.

Or everyone may begin making that tiny orchestral movement toward their water bottles, which means the experience has concluded spiritually even if the conductor has not noticed.

Cull.

After a month of this, I may have nothing. That seems entirely possible. One thousand symphonies, or the genetic remains of 1,000 symphonies, scattered across a hard drive.

Nothing great. Nothing good. Just a magnificent quantity of proof that abundance does not manufacture greatness.

Or, I may find one peculiar thing. One. Something that survives because every time I nearly throw it away, it does something else. Something I keep wanting to hear. Something musicians want to play again. Something that statistically should have disappeared somewhere around Round Four but did not.

And because I am using 4.5, there is another pleasing possibility.

By the time I know what I found, the machine that made it may no longer exist.

Suno will have moved on. The better models will be better. The audio will be cleaner. The controls will be finer. The future will have arrived on schedule. And I will be sitting there with Symphony No. 683, made in four days by a deprecated computer program no one is allowed to use anymore.

This is where the whole thing begins to acquire the rancid sweetness of nostalgia.

We may someday look back on Suno 4.5 the way Star Trek fans look back on the late 1960s. Three seasons. That was it. Three seasons, and then someone in a room looked at the thing and said, no, thank you, we have enough of that!

Only here it may be worse.

At least Star Trek remained visible. You could rewatch it. You could build conventions around it. You could dress as Spock and make your spouse reassess several decisions.

But Suno 4.5 is not merely being canceled.

It is an instrument being taken apart.

You will still be able to hear what was made with it, but no one will be able to make new things in quite that way again. The plug pulled. The animal gone. The little patch of musical wilderness paved over for a cleaner model with better retention, sharper vocals, more reliable structure, and fewer moments where the low brass suddenly appears carrying a knife.

This may make the survivors artificially scarce.

Not valuable in the normal collector sense. I am not imagining Sotheby’s auctioning Symphony No. 683 while a man in a scarf nods gravely into a telephone. Though now that I have said it, I am imagining exactly that.

No, I mean historically scarce.

There may come a time when people listen back to early AI music and say, wait, they had this? This strange, unstable, overproductive, genre-drunk machine that could accidentally wander into a musical dialect nobody asked for?

And they thought it sucked?

And then they got rid of it?!

What were they thinking?

Which is, of course, what people always say later. About orchestras. About recordings. About cities. About restaurants. About ugly buildings that became beloved once nobody could afford to build anything with that much personality again.

The future has a genius for turning yesterday’s defects into today’s lost virtues.

Tape hiss becomes warmth. Bad color correction becomes period texture. A canceled television show becomes scripture.

And perhaps one day, some student in 2077 will be assigned an early AI symphony made on Suno 4.5 and hear, underneath the underwater bassoons and synthetic cymbal dandruff, something nobody can quite reproduce anymore.

Not because it was perfect. But because it wasn’t. The machine still wandered. It had not yet been optimized into obedience.

And somewhere in the fossil record there were things called “Chain Gang Stravinsky” and “I’m in Love with the Chickens,” and people were expected to go on with their day after hearing them.

Because, for one brief and ridiculous moment, before the engineers cleaned the zoo, the animals were loose.

Copyright © 2026 by Paul Henry Smith

Let me show you what I’m talking about.

Imagine a small appliance sitting on the counter in a pleasant office kitchen, the sort of place with pale wood, expensive plants, and a bowl containing three pieces of fruit that nobody has touched in weeks. The machine is cream-colored and reassuringly solid, with rounded corners and a large black screen. It looks like something Braun might have made in 1973 if Dieter Rams had been given access to a molecular-generation model and told that the future depended on snacks. On the screen is an enormous three-dimensional cloud of tiny luminous points. Some gather into dense yellow islands, others into green continents or purple filaments. There are labels here and there: CITRUS, GREEN LEAF, FERMENTED, FLORAL, ROASTED. But the most interesting thing on the screen is not what has been labeled. It is everything that hasn’t.

You move the cursor away from the green cluster and into the darkness between GREEN LEAF and FERMENTED, stopping in a sparse region that belongs to neither. There is nothing there because, as far as anyone knows, there is no flavor there. You press the mouse button and hold it. A circle expands around the cursor. New points begin flickering into existence. The machine is not searching a catalog for something that tastes a bit like cucumber. It is generating molecular structures predicted to create a sensory experience in that particular unoccupied territory. Candidates appear and disappear. One is unstable. Another would be miserable to synthesize. Another trips an early safety screen. Eventually a candidate remains, and much later—after considerably more chemistry, characterization, toxicology, regulatory work, and controlled testing than the tasteful progress indicator on the screen would lead you to imagine—a little ledge slides out of the machine. On it sits a transparent bubble containing a single clear droplet. A narrow paper flag curls out of the top, like the plume on a Hershey’s Kiss, and on it are the words LUMEN-7A62: green mineral · airy · crisp. Nobody has ever tasted it before, for the excellent reason that until somebody asked the machine to look in that particular patch of darkness, it had never existed.

I keep returning to this image because it makes visible a mistake we have been making for several thousand years, albeit a productive and often delicious one. We have confused flavor with ingredients. For practically the whole history of cuisine, flavor has been downstream of biology. We have strawberries, garlic, cows, cacao, limes, mushrooms, chickens, coffee beans, and an almost vindictive number of peppers, and so we have become very clever about persuading these things to do tricks. We roast them and ferment them, dry them and smoke them, breed them, age them, distill them, bury them, combine them, and occasionally allow microorganisms to have their way with them for six months before announcing that the smell is intentional. The results have been magnificent. I have no complaint against butter. But underneath this entire culinary civilization sits an assumption that is beginning to look less like a law of nature than a historical inconvenience: flavor comes from ingredients.

An ingredient, after all, is just one way of producing a sensory event in a human being. A strawberry is an extraordinarily elaborate biological machine for delivering a particular collection of molecules to your nose and tongue, along with water, sugar, acids, seeds, fiber, color, nostalgia, and the occasional disappointing white interior. We have treated the strawberry as the fundamental object because, until recently, there was not much practical reason to do otherwise. But if the thing we actually care about is the experience produced when those molecules meet the human sensory system, then the strawberry begins to look less like the definition of strawberry flavor and more like one implementation of it. Once that distinction becomes clear, a peculiar door opens. We can stop asking only what nature has given us to taste and start asking what the human sensory system is capable of experiencing.

Flavor science has traditionally approached this from the sensible direction. Take a molecule, expose someone or something to it, and determine what it does. Molecular structure goes in; perception comes out. Increasingly, machine-learning systems can participate in this process, predicting whether a compound is likely to be bitter or sweet, estimating odor character or intensity, and learning relationships between molecular structure and sensory descriptors. Generative systems have also begun proposing novel odorants and taste-active molecules, including candidates that can subsequently be synthesized and experimentally evaluated. None of this requires the invention of magical technology. The strange thing is that so many of the pieces already exist. The more interesting move is simply to reverse the arrow.

Instead of asking what a molecule will taste like, ask what molecule would taste like this. Perhaps I want something with the opening brightness of yuzu, some of the vegetal snap of a tomato leaf, a peculiar mineral sensation in the middle, almost no sweetness, and an aftertaste that vanishes completely after eight seconds. That description can become a target in a multidimensional sensory space. A generative model can then search chemical space for structures predicted to produce it, while other models reject candidates that appear unstable, impractical, reactive, environmentally troublesome, or otherwise unpromising. What sounds at first like an eccentric application of artificial intelligence is really an inverse-design problem, of the sort appearing across materials science, drug discovery, and protein engineering. We know approximately what behavior we want. Instead of waiting to encounter something that exhibits it, we generate structures that might.

The interface for such a system should not, I think, look much like chemistry software. It should look like an instrument. I call it FlavoSynth, partly because it describes what the machine does and partly because every sufficiently interesting technology eventually deserves a name that would have looked good stencilled on a synthesizer in 1982. A musician does not sit at a synthesizer thinking about Fourier transforms. The complexity has been translated into controls corresponding to perceptual consequences: attack, decay, resonance, brightness. FlavoSynth could do the same for flavor. You might turn up brightness, pull back warmth, add greenness, lengthen persistence, or, most important, rotate a large knob marked STRANGENESS. The chemistry remains underneath, where chemistry is happiest. The person using the machine manipulates sensation.

There is another version of this idea that feels less like playing an instrument and more like committing tasteful acts of vandalism. I call it FlavorShop. Load strawberry—not a photograph of a strawberry but a representation of its sensory character—and begin editing. Increase brightness by eighteen per cent, reduce jamminess by twelve, introduce a little more green, leave sweetness exactly where it is, and stretch the experience so that it persists for nine seconds rather than four.

The familiar tools of image editing become strangely natural when transferred to perception. An eyedropper could sample the cold mineral character of cucumber so that you could paint some of it into watermelon. An eraser could remove the faint cooked note from a processed fruit flavor. Layers might separate aroma, taste, trigeminal sensation, texture, and temporal behavior. A magic wand could select whatever combination of signals humans interpret as “ripe,” which is the sort of phrase that sounds ridiculous until you realize that computers already perform similarly improbable acts with images every day.

Eventually, of course, somebody would add Generative Fill. You would circle an empty region in the sensory representation and type, “Put something here that we have never tasted before.” At that moment FlavorShop would cease to be a sophisticated way of improving strawberry yogurt and become something considerably stranger. It would be an editor for possible human experience.

To make such an editor useful, we would need something like a Flavor Atlas, a map of known sensory chemistry. Imagine every characterized flavor-active molecule rendered as a point in an enormous space, positioned according to what it does perceptually rather than where chemists happen to place it taxonomically. Familiar territories would emerge as dense constellations: citrus, floral, roasted, green, fermented, sulfurous, marine. The shape of the map would be fascinating, but the revelation would come from its emptiness. Between the known clusters would be enormous gaps, regions of molecular and perceptual possibility that cuisine has never had any reason to visit. Our present vocabulary might prove to be less a description of flavor itself than a record of the neighborhoods biology happened to build.

This is where our language starts to become a handicap. When we imagine an unfamiliar flavor, we tend to construct it from familiar ones: pineapple crossed with basil, perhaps, with a little sea air. But that may be like trying to describe an unknown color by listing vegetables. Fruity, floral, woody, nutty, smoky, green, meaty—these categories exist because we have repeatedly encountered things that produced those sensations. If there are other stable and pleasurable regions of sensory space, we should not expect to have words for them in advance. The first encounter with a genuinely unfamiliar flavor may not produce a lyrical tasting note. It may produce a person staring at a tiny transparent bubble and saying, with commendable scientific precision, “What the hell is that?” The name can come later.

We might also discover that the very idea of a flavor is too static. Eating takes place through time, but most flavor descriptions flatten the experience into a noun: cherry, vanilla, smoke. A designed flavor could instead be treated as a trajectory. It might begin with a sharp green attack, reveal tropical sweetness two seconds later, develop an unfamiliar cool mineral quality around the fourth second, lose its fruit at seven, and disappear by ten. Perhaps one molecule could produce such behavior through its volatility and receptor interactions; perhaps the better solution would be a tiny ensemble whose members arrive and recede in sequence. These would be flavor chords, compositions whose identity exists not merely in their components but in the order and duration with which they are perceived. Once flavor acquires attack, development, counterpoint, and resolution, cuisine begins to edge toward music.

All of this naturally suggests a business in novel flavor molecules, and there will presumably be one. But I suspect that selling the molecules is not where the deepest value accumulates. The valuable thing is the record created while discovering them. Existing flavor datasets are small and awkward compared with the complexity of what they attempt to describe, because flavor does not reduce neatly to a molecule and an adjective. Concentration matters. Temperature matters. Fat, water, alcohol, pH, starch, and neighboring compounds matter. Time matters. Humans matter, which is almost always where an otherwise tidy scientific problem becomes expensive.

The company I would want to build, therefore, would not merely generate molecules. It would generate experiments. The model proposes a candidate; the candidate is synthesized and characterized; its physical properties are measured; appropriate receptor, instrumental, toxicological, and eventually sensory evaluations are performed; and the result returns to the model. Each pass through the loop adds another precise correspondence between chemical structure and human perception. After ten thousand such experiments, you possess something unusual. After a hundred thousand, you possess something competitors cannot reproduce by downloading the latest model weights from GitHub. You have a proprietary map connecting matter to subjective experience, including the failures, the concentration effects, the strange interactions, the temporal curves, and the molecules that everyone expected to smell wonderful but instead reminded six out of eight panelists of a damp vacuum cleaner.

That dataset becomes the flywheel. Better experimental data improves the perception model, which improves the generated candidates, which makes synthesis and testing more productive, which creates still better data. Model architectures will spread; techniques will be published; generative chemistry itself will become increasingly commonplace. But the accumulated physical history—we made this molecule, at this purity and concentration, placed it in this matrix, and this is what people experienced over the following twelve seconds—is difficult to copy. The company that sets out to discover new flavors may eventually discover that its most valuable asset is not any particular flavor at all. It is the atlas.

So this is an exhortation to the people already standing closest to the machinery: flavor chemists, sensory scientists, computational chemists, food scientists, olfaction researchers, molecular-generation people, robotic-lab people, and the flavor companies that already synthesize large numbers of candidate compounds every year. Generative AI should not merely be used to make the existing search slightly faster, or to find a cheaper route to a familiar strawberry note. The more interesting opportunity is to change what is being searched for. Build perceptual spaces. Invert them. Generate toward sensory targets. Map the empty regions. Model flavor through time. Reward novelty without confusing novelty with safety. Create instruments that let flavorists manipulate perception directly, then connect those instruments to disciplined physical experimentation.

We have spent thousands of years exploring the flavors that Earth’s biology happened to make available, and there is something charming about the possibility that we have mistaken this local abundance for completeness. Evolution was not conducting a systematic search for deliciousness. It was solving problems involving reproduction, predation, energy storage, pollination, defense, and survival, after which we arrived with napkins and developed strong opinions about mouthfeel. Coffee deserves every bit of its cultural status. Garlic has performed heroically. Chocolate has nothing to apologize for. But none of these facts implies that the molecular possibilities capable of delighting a human nose and tongue have been remotely exhausted.

Somewhere in chemical space there may be a sensation as distinctive as mint, smoke, coffee, or chocolate that no human being has ever experienced. There may be thousands of them. We cannot describe them, because description has to wait for encounter, and encounter has so far been constrained by what plants, animals, microbes, heat, time, and a few centuries of industrial chemistry happened to put in our path. Generative chemistry gives us a way to go looking deliberately. We can press into the blank parts of the map, generate candidates, reject almost all of them, investigate the survivors carefully, and gradually discover whether the boundaries of our cuisine have been the boundaries of perception or merely the boundaries of our ingredients.

I strongly suspect the latter. It would be an astonishing coincidence if, among the immense number of molecular structures capable of interacting with human sensory systems, evolution and agriculture had already stumbled across all the pleasurable ones. The universe is under no obligation to have made its entire menu available in the produce section. We have finally acquired tools that might allow us to ask what else is there, and it seems almost rude not to ask. Surely, after several billion years of chemistry and a few hundred thousand years of human consciousness, we have not reached the absolute zenith of flavor. Surely the universe has something left to say. It simply cannot be that the final answer was ranch.

Copyright © 2026 by Paul Henry Smith

I am looking at the Copyright Office’s recent report on artificial intelligence and copyrightability, trying to figure out a fairly basic question: when I use a generative music system as part of composing a piece, at what point does the Copyright Office think I stop being the composer? The report is actually rather reassuring at first. It does not say that using AI makes a work uncopyrightable. Quite the opposite. It says AI can assist human creativity, that human-authored material appearing in an AI-assisted work remains protectable, that creative selection, coordination, arrangement and modification can count, and that these things have to be examined case by case. Good. This sounds sensible.

Then I encounter the test the Office wants to use.

We are supposed to ask whether the computer was merely an assisting instrument or whether the “traditional elements of authorship,” including musical expression, selection and arrangement, were actually conceived and executed by the machine. And here is the part that catches my attention: the Copyright Office did not invent this language for ChatGPT. It is quoting itself from 1965. (U.S. Copyright Office)

Traditional elements of authorship.

Well, all right. Music has traditions coming out its ears. I know a little about those. And 1965 is an interesting date. Computers were not exactly banging out Spotify singles between lunch and dinner. What on earth was happening in 1965 that caused the Copyright Office to be thinking about whether a machine could make musical expression? Now I’m curious. So I go back and look.

And there it is!

The Copyright Office’s 1965 annual report says it had already received an application for “a musical composition created by computer.” Not contemplated receiving one. Not worried that someday an IBM mainframe might get ideas above its station. It had one. Sitting there. Somebody had already walked through the door with computer-generated music and said: copyright this. The Office immediately saw the problem we are supposedly confronting for the first time today. Is the computer simply an instrument used by a human author, or did the machine itself conceive and execute the musical expression, selection and arrangement? (U.S. Copyright Office)

Now I really want to know what happened.

Because if the Copyright Office is reaching back sixty years to tell me that its traditional framework already knows how to deal with AI, I would quite like to see what it actually did when the tradition was being formed.

Enter Lejaren Hiller and Leonard Isaacson.

In the 1950s, Hiller and Isaacson used the ILLIAC computer at the University of Illinois to generate musical material for the Illiac Suite. This is not MIDI playback. The computer is participating in composition. Hiller and Isaacson establish musical rules and constraints, the computer produces pitches and rhythms under those conditions, chance is deliberately involved, and the resulting material becomes a string quartet.

The score is published in 1957, its foreword makes the computer’s role quite clear, and it receives copyright registration EP110997. According to recent archival scholarship, Hiller and Isaacson deliberately resisted simply cherry-picking whichever machine outputs sounded prettiest. Their human arranging was comparatively limited: they established the larger formal structure, chose tempi and made a few practical transpositions. The computer was not hiding behind the curtain pretending to be a pencil. It was doing exactly the thing that now makes everybody nervous. It was making musical choices. (IU Robert H. McKinney School of Law)

Okay. That is interesting.

Then Hiller develops MUSICOMP and makes Computer Cantata. The system lets a composer enter musical parameters and instructions that govern the generation of a composition. Some of the piece evolves probabilistically.

The archival record says the work was submitted to the Copyright Office in December 1963 and ultimately registered as EU0000804754, later renewed. The same scholarship makes a rather persuasive case that Computer Cantata is almost certainly the computer-created composition the Copyright Office was referring to when it wrote that famous paragraph in 1965. (IU Robert H. McKinney School of Law)

Wait, what?

So the Copyright Office today points to 1965 and says, in effect, we have been asking the right question all along. But the very historical moment it invokes contains computer-generated pitches, probabilistic compositional procedures, systems that receive musical instructions and elaborate them, and registrations issued for the resulting compositions.

This is getting good.

Suddenly the question is no longer whether AI represents some alien incursion into composition. We already have a seventy-year tradition of composers building systems that make musical decisions for them. And once you start pulling on that thread, it gets worse, or better, depending on how fond you are of intellectual trouble.

John Cage wanders in.

(Of course he does.)

Cage is a particularly irritating witness for anyone who thinks authorship means personally deciding every event, because Cage sometimes made an extraordinarily sophisticated artistic decision not to make particular decisions.

Chance was not an accident that leaked into his music because his technology was inadequate. Relinquishing control was part of the composition. He could decide what variables existed, what possibilities were permitted, how choices were made, when they happened, what remained fixed and what was allowed to float. His Music of Changes, built around chance operations, was accepted by the Copyright Office as a published musical work in 1961. (IU Robert H. McKinney School of Law)

Then in 1977 the Copyright Office gets nervous about Cage’s Renga, and this becomes almost absurdly relevant. Renga uses hundreds of graphic images that performers interpret. The Office had even heard a performance. Its examiners could tell there was musical authorship happening somewhere, but they were having trouble working out how much belonged to Cage and how much belonged to the orchestra. So they ask. Cage responds, essentially, that if the orchestra did not have the materials he created, it would not produce this music; give those same materials to another orchestra and the details would change, but the work would remain recognizably itself. That satisfies the Office. Renga is registered. (IU Robert H. McKinney School of Law)

I love this, because Cage is not saying, Look, I secretly specified all the notes.

He is saying something far more interesting:

I authored the musical circumstances in which those notes happen.

That is a very different account of composition.

And now I go back to the Copyright Office’s AI report with this rattling around in my head, and I find the phrase that nearly makes the whole thing fall over.

The Office now considers iterative prompting. Suppose somebody generates something, inspects it, changes the instructions, generates again, listens, changes them again, perhaps does this dozens or hundreds of times, and eventually reaches a result. The Office is skeptical that this increases the user’s authorship of the resulting expression.

Changing and resubmitting the prompt, it says, is basically “re-rolling the dice.” You get more candidates to choose from, but you have not increased your control over how the system turns the instructions into expressive details. (U.S. Copyright Office)

I’m sorry. Re-rolling the dice?

In music??

This may be one of the most unfortunate metaphors the Copyright Office could possibly have chosen.

Composers have literally rolled the dice.

They have tossed coins. They have used the I Ching. They have written probability tables. They have built stochastic systems. They have created algorithms whose entire purpose was to make decisions that the composer deliberately chose not to make directly. The dice are not necessarily evidence that nobody is composing. Sometimes deciding what the dice mean is the composition.

That is bananas!

Now, I don’t want to make the opposite mistake. Suppose I type “write a really exciting piano concerto” into a box, receive twelve concertos, pick number seven because the ending gives me goosebumps, and then announce that I am Sergei Rachmaninoff. I understand the Copyright Office’s problem. Selecting something because I like it is not necessarily the same as authoring it.

Hard work is not authorship either. One may spend a fantastically long time pulling the lever of a slot machine without designing the cherries. The Office is right to resist a theory in which persistence alone turns generation into composition. Its report explicitly says that current general-purpose prompting, by itself, ordinarily does not provide sufficient control, while leaving protection available for human inputs, arrangement and modification. (U.S. Copyright Office)

Fine.

But now we have to find the boundary.

Because Cage exists. Hiller exists. Algorithmic composition exists.

And the musical act of deliberately deciding how a decision will be made cannot simply disappear from authorship because the mechanism doing the deciding got fancier.

At this point I can already hear somebody saying: Cage, Hiller, stochastic music, experimental computer laboratories in Illinois, yes yes, very clever. The twentieth century produced a number of nut-job composers whose principal contribution to civilization was causing perfectly nice people to flee auditoriums. Are we really going to construct a theory of copyright around them?

Fair enough.

Forget Cage.

Let’s talk about John Williams.

And this may be the example that blows the argument open, because it takes us out of the experimental-music corner entirely.

Nobody seriously wonders whether John Williams is a composer. He wins the Oscar for Star Wars. The Academy record says exactly that: Music, Original Score, winner John Williams. Williams gets up and thanks George Lucas, Gary Kurtz, the London Symphony Orchestra, recording engineer Eric Tomlinson and, interestingly, Herb Spencer. Spencer was one of Williams’s orchestrators; the Library of Congress, for example, separately credits Williams as composer and Herbert Spencer as orchestrator on Empire of the Sun. (AAS Speeches Database)

Now stop there.

Orchestrator is an actual job.

Why?

Because a composer can create a musical work without personally specifying every musical detail required to produce its final orchestral realization.

Depending on the composer and the project, an orchestrator may make decisions about instrumental distribution, register, doublings, voicings and other aspects of realizing music for the orchestra. The Copyright Office itself recognizes that new harmony or instrumentation can constitute authorship in a musical arrangement. (U.S. Copyright Office) These are not clerical details. Move a chord from four trombones to divided violas and clarinets and you have not changed the billing address. You have changed the music.

Yet nobody examines a John Williams score, finds an expressive choice made by an orchestrator and gasps:

My God. The composer has vanished.

There is no emergency hearing at the Academy.

Nobody repossesses the Oscar.

Somehow, without difficulty, we understand that musical authorship can be hierarchical. One person may originate and direct the musical conception while another person makes additional expressive choices in realizing it.

And I suddenly realize what an orchestrator looks like if you describe the job in deliberately mechanical language.

A composer supplies a highly compressed musical representation. Melody. Harmony. Rhythm. Form. Character. Dramatic function. Some instrumental information. Perhaps a great deal more. The orchestrator receives this specification, combines it with a huge internal model of orchestral practice, understands relationships not exhaustively spelled out in the input, and produces a much more detailed realization containing hundreds or thousands of local musical decisions.

High-level musical instructions go in.

Detailed musical realization comes out.

Excuse me.

Have I just described an orchestrator or a generative model?

Obviously the analogy is imperfect. The orchestrator is a person. The orchestrator can be an author. The orchestrator has judgment, agency, experience, taste, contractual relationships and possibly strong opinions about division cellos. Neural networks do not become people because they can voice horns nicely. I am not trying to smuggle AI personhood into copyright by putting Herb Spencer glasses on a GPU.

But that does not rescue the proposition we are examining.

If the rule is:

You did not choose the expressive details, therefore you did not compose them,

then human orchestrators have been blowing holes through that rule for generations.

They choose expressive details.

The composer remains the composer.

So something else must be going on.

And now the whole essay changes. I am no longer primarily asking whether the Copyright Office is being unfair to Cage. I am asking something much more ordinary and dangerous:

How granular does a composer have to be?

Every note?

Really?

Consider a lead sheet. Berklee describes its pure form as melody plus chord symbols and specifically notes that rhythm-section players frequently develop their own parts. The keyboard player voices chords. The drummer makes fills. The bassist invents a bass line. Harmony, texture, voicing and orchestration may not be specified on the page at all. Yet we have no difficulty recognizing the underlying song as a composition. (Berklee)

Go back a few centuries. Figured bass provides a bass line and harmonic information while allowing the player freedom in choosing the actual chord voicings. A Baroque keyboardist may realize the same harmonic framework differently from another player. (Music Theory at Puget Sound)

Who chose every sounding note?

Not necessarily the composer.

Do we have a problem?

Apparently not.

Now zoom in further. A violinist sees a sustained note. Did the composer specify the vibrato rate? Width? Exact moment the vibrato begins? Bow speed? Contact point? Minute fluctuations of intonation? The precise shape of the attack? Those are expressive musical decisions. Performers can make one note ache, glow, bark, hover or go completely dead without changing its pitch or nominal duration.

Has the composer failed the expressive-elements test because the violinist supplied them?

What if the piece uses a rain stick (an instrument containing beads or seeds tumbling through a resonating chamber)? Does the score need to identify where seed number 418 hits the interior wall at 12.4 seconds? If it does not, who composed that bit?

Gravity?

Clearly at some point we have put the microscope at the wrong magnification.

A musical composition has never been a lossless specification for every vibration of air in every valid performance. Music works because authorship can exist at different levels of abstraction. Sometimes the composer specifies the pitch. Sometimes the chord. Sometimes the relationship between chords. Sometimes a transformation. Sometimes a pattern. Sometimes a set of options. Sometimes the direction in which another musician should exercise judgment.

And once that thought lands, Cage stops looking quite so eccentric.

He is just taking an ordinary feature of music and making it explicit.

The violin concerto already leaves things open. Cage composes the openness.

And that phrase keeps getting more important:

Cage was not failing to compose the decisions he left open. He was composing their openness.

That is a richly artistic choice. It is also a richly intellectual one. What happens if the work does not consist of a single predetermined path through musical time but instead defines a family of paths? What happens if the composer deliberately moves authorship upward, from specifying each event to specifying the conditions under which events may occur?

By definition that is not less composition. It is composition taking place somewhere else.

And just when I think I have enough trouble, Bach arrives.

(Of course, Bach arrives.)

I had originally been thinking of the Goldberg Variations as a fairly gentle example of higher-order compositional thinking. The work is famously organized. Every third variation is a canon, the interval of imitation rises systematically through the sequence, and the whole structure is saturated with relationships that exist above the level of individual notes. Fine. Bach creates notes, obviously, but he also builds machinery among them. (Bach Museum Leipzig)

Then I discover there is more.

Bach’s own personal copy of the published Goldberg Variations was found to contain, handwritten on a blank page, fourteen additional canons built on the first eight fundamental bass notes of the Aria. The Bach Museum describes exactly this discovery. (Bach Museum Leipzig)

Eight notes.

Bach starts subjecting them to operations.

Now the algorithmic-composer alarm begins blinking.

I want to be careful here. These fourteen canons do not prove that Bach had some eighteenth-century goldberg.generate() function that manufactured the original thirty variations while he went off for a beer. But scholarship on BWV 1087 shows that the tiny eight-note object is extraordinarily generative.

Matthew Haakenson’s analysis describes the subject as structurally predisposed to generate canonic material and develops methods for replicating several of the canonic types. Peter Harrow goes after Canon 10 and finds that Bach has supplied three procedures: double counterpoint, melodic inversion and vertical shifting. Combine those and Harrow gets 56 possibilities. Then he adds another perfectly natural variable, horizontal displacement of one voice against the other.

The number becomes 840.

Eight notes.

Eight hundred and forty apparent possibilities from this modest enlargement of the machinery, and Harrow notes that augmentation, diminution, retrograde and other procedures cause the possibility space to grow further. (College Music Symposium)

At this point I am grinning.

Because this is not the romantic picture of composition in which genius sits alone at a candlelit desk and lovingly decides E-flat, then perhaps pauses for a spiritual experience before deciding F.

This is Bach designing a musical space.

He specifies material. He specifies transformations. He discovers what survives them. He constructs procedures capable of yielding families of musically meaningful results.

The artistic intelligence is not located only in whichever particular notes eventually appear on the page. The algorithm can be where the composer put some of it.

Now go back to AI.

Suppose I supply a motif and say: construct an atonal scherzo from this. I want a radically contrasting texture every two bars. I want countermelodies derived from the motif, but genuinely independent, not limp doublings. I want phrase endings to taper in rhythmic activity and density. The model gives me something. No, the counterpoint is tracking the melody too closely. Push it into contrary motion. Again. Better. Keep this passage. Throw out that one. Change the relation between these sections. The return is happening too early. Reshape it. Preserve the original motif there. Now make the texture collapse rather than merely thin. Eventually I have a piece. I notate it.

What exactly happened?

The easy answer is: the AI chose many pitches and rhythms.

Yes.

So did Hiller’s machinery.

So can an aleatoric process.

So can an orchestrator.

So can a continuo player.

So can the pianist looking at C7 on a lead sheet.

That fact alone does not answer the authorship question.

Now I want to know what the Copyright Office means by control.

Suppose I say, “write an atonal scherzo.”

Very broad.

Then I supply the motif.

More control.

Then I determine its formal divisions.

More.

Then the transformations.

More.

Then the contrapuntal behavior.

Then the location and nature of textural contrast.

Then phrase-shape rules.

Then constraints on register.

Then I reject realizations that violate the relationships and redirect the system.

At what instruction do I become the composer?

Is there a little copyright light that turns green?

More disturbingly, consider the question in reverse.

At what increase in the machine’s competence do I stop being the composer?

That is where I think we are getting close to something genuinely strange.

Imagine I build a primitive algorithm myself. I specify a pitch set, probabilities, transformation tables, counterpoint rules, rhythmic distributions and formal logic. I write a lot of code. The system eventually chooses individual pitches. Everyone looks at this and says: algorithmic composition. This has been a recognized musical practice for decades.

But now, improve the machine.

Instead of writing fifty lines of code describing how rhythmic density should diminish toward a cadence, I can tell the system, “taper the rhythmic density toward the end of the phrase.”

Same musical intention. Better instrument.

Have I become less of a composer because my instrument understands me?

Improve it again. Now I can express the musical relationship more abstractly still. The machine knows what a scherzo is. It understands contrary motion. It has representations of orchestral texture. It can infer what I mean by making one passage feel suddenly skeletal.

At what point does the machine become so good at realizing musical instructions that my ability to give musical instructions ceases to count as composing?

If this is the principle, then we have arrived somewhere very odd:

A composer may delegate musical decisions to a machine only until the machine becomes sufficiently good at making musical decisions.

Bananas again?

Try that with the Hollywood orchestrator. I hire an inexperienced orchestrator who requires me to specify every doubling, every voicing and practically every instrument. Fine. Clearly I am the composer.

Then I hire someone brilliant.

Now I can say less because the orchestrator understands more. I can express the musical objective at a higher level and the orchestrator produces the detailed realization I had in mind, perhaps adding solutions I would never have specified individually.

Did my authorship decrease because the assistant became more capable?

Did John Williams become less John Williams when Herb Spencer understood what he wanted?

Nobody thinks that way.

Yet with AI we are in danger of constructing exactly this weird inverse relationship between the intelligence of the tool and the authorship of the human. The better the instrument becomes at realizing intention, the less intention apparently matters. Carry that far enough and the only composer whose copyright is completely safe will be the one prudent enough to use sufficiently stupid tools.

That would be an astonishing definition of progress.

Now I go back again to the Copyright Office’s phrase: traditional elements of authorship.

Traditional.

This is where I think their own word begins closing the net.

If the Office really means “traditional,” then it cannot mean an imaginary tradition in which the composer personally selects every expressive atom. That tradition does not exist.

The actual tradition is Bach writing musical machinery.

It is Baroque composers providing harmonic blueprints that performers realize.

It is jazz and popular composers writing lead sheets from which musicians invent specific parts.

It is the film composer handing material to an orchestrator.

It is every classical performer making thousands of expressive decisions no score could possibly encode.

It is Cage deliberately moving decisions outside himself.

It is Hiller writing programs that generate pitches and rhythms.

It is algorithmic composers spending seventy years deciding not merely which notes to choose, but how notes will be chosen.

And, crucially, the Copyright Office itself has not merely observed this history from a safe distance. It was there. It actually registered Illiac Suite. It registered Computer Cantata. It wrestled directly with Renga and then accepted Cage’s explanation that the details could change while the work remained recognizably his. Its own 1965 report refused to say that use of a computer categorically disqualifies authorship and demanded attention to the actual circumstances. (IU Robert H. McKinney School of Law)

So now when the modern Copyright Office tells me that prompting is insufficient because the model is choosing the expressive elements, I don’t want to dismiss the concern. I want to ask them to keep going.

Which expressive elements?

At what level?

How much delegated realization is too much?

Why does delegation to an orchestrator leave the composer standing, while delegation to an algorithm potentially removes authorship from the resulting expression?

Why does a probability procedure remain part of a composer’s technique until the procedure is implemented by a sufficiently powerful neural network?

Why is unforeseeability evidence against authorship when composers have deliberately cultivated unforeseeability as an expressive condition?

Those are not rhetorical gotchas. There may be good answers.

Generative AI really is different from figured bass. A model can respond to an absurdly vague request with several minutes of astonishingly detailed music. The causal relationship between instruction and result can be opaque even to the people who built the system. The Copyright Office is absolutely right to resist the notion that anybody who types a sentence into a music generator becomes the author of everything that tumbles out. Its report notes that identical prompts can yield differing results and that current systems may introduce substantial content never specified by the user. That genuinely complicates the relationship between conception and execution. (U.S. Copyright Office)

But “complicated” is not the same as “unprecedented.”

And complexity is not a license to collapse the hierarchy of musical authorship into a note-counting exercise.

The interesting question is not whether the machine made decisions. Of course it did.

The question is what the composer decided about those decisions.

Did the human provide the musical materials? Establish the transformations? Determine the invariants? Define the relationships? Decide which dimensions may vary? Shape the larger form? Diagnose failures? Alter the governing instructions? Preserve or discard realizations according to a musical conception? Fix a particular realization as the work?

At some point those things can plainly be too weak to constitute authorship.

At some other point they can plainly be composition.

And between them lies the territory we actually need to understand.

This is why that seemingly innocent “re-rolling the dice” line bothers me so much. It carries within it an assumption about what the human is doing. It pictures the user outside the creative system, waiting for luck to deliver something appealing.

But the composer may have designed the dice.

The composer may have decided what is written on each face.

The composer may determine which dice are rolled at which moments, which outcomes are legal, how one result alters the next roll, when the rules change, what musical material is exposed to the procedure and when the procedure must stop.

At that point, telling me that the composer did not personally choose the number four is missing the composition.

The dice are not the composer.

But neither are they proof that there is no composer.

They are an instrument.

And this is where, after going around this peculiar historical loop, I find myself agreeing with the Copyright Office more than I expected.

Maybe generative AI doesn’t require some exotic new theory of authorship.

Maybe the old one really is enough.

But then we have to bring the old one with us.

All of it.

Not a nostalgic caricature of the composer as a solitary person who manually injects every note into the score, but the actual musical tradition, full of delegation, realization, indeterminacy, performers, orchestrators, algorithms, procedures, shorthand, open variables, chance and intelligent systems for making choices.

The Copyright Office itself tells us to look for traditional authorship.

Wonderful. Let’s. Because once we do, the exceptions begin piling up until they stop looking like exceptions.

Maybe composition has never meant merely choosing the notes.

Sometimes it means choosing the notes.

Sometimes it means choosing the relationships among the notes.

Sometimes it means choosing the rules governing those relationships.

Sometimes it means creating the material and telling another musician how to elaborate it.

Sometimes it means defining a space inside which another intelligence gets to move.

Sometimes it means deliberately refusing to choose a detail because the openness itself is part of the work.

Cage was not failing to compose the decisions he left open.

He was composing their openness.

Bach was not surrendering intelligence when he devised transformations capable of producing hundreds of contrapuntal possibilities.

He put some of the intelligence into the system.

John Williams does not cease to be the composer because an orchestrator intelligently realizes musical details Williams did not personally specify.

The intelligence of the realization and the authorship of the composition are not mutually exclusive.

So if neural networks really overturn this long arrangement, fine. But now the burden shifts. The Copyright Office cannot merely tell us that the machine chose expressive elements. It has to tell us why that particular kind of delegated choice is different from all the other delegated choice music has absorbed without losing the composer.

Why does the orchestrator not erase John Williams? Why does the continuo player not erase Bach? Why does the jazz pianist not erase the songwriter? Why did chance not erase Cage? Why did computer-selected pitches not erase Hiller? Why does a composer remain the author when another human being elaborates a compressed musical instruction into thousands of detailed choices, but become suddenly suspect when a machine becomes capable of doing something structurally similar?

There may be an answer. Neural networks are not orchestrators. They are not performers. They are not dice, or figured bass, or an IBM 7094. Their powers are different enough that perhaps copyright really does need to draw a new line somewhere. But if so, where is the line? It cannot simply be where the composer stops choosing individual notes, because music crossed that line centuries ago.

It cannot be where another intelligence contributes expression, because Hollywood crosses that line every afternoon. It cannot be where the result becomes unpredictable, because composers have made unpredictability into an artistic medium. And it surely cannot be where the tool becomes capable enough to understand the composer’s intentions at a higher level, because that would leave us with the bizarre rule that musicians preserve authorship only by using sufficiently stupid instruments.

This is why I keep coming back to the Copyright Office’s own word: traditional. It wants traditional authorship. Wonderful. But the actual tradition is not a fellow hunched over manuscript paper manually selecting every expressive atom. The actual tradition is full of delegation, shorthand, realization, chance, open variables, orchestration, algorithms, performers and systems. Cage was not failing to compose the decisions he left open. He was composing their openness. Bach could place compositional intelligence inside a set of transformations. John Williams does not cease to be the composer because somebody else intelligently realizes details he never wrote down.

And perhaps that is the thing generative AI has exposed. We thought we knew what a composer was because, most of the time, nobody forced us to say exactly where composition stopped and realization began. The distinction could remain pleasantly fuzzy. The orchestrator went to work, the violinist added vibrato, the continuo player supplied notes that were not on the page, the algorithm rolled its dice, and somehow civilization continued without requiring us to perform an authorship autopsy on every measure.

Now the machine has become good enough that the fuzziness can no longer be ignored. It can operate almost anywhere along the chain, from obedient pencil to highly capable musical elaborator. And suddenly we have to decide whether authorship belongs only to the person who makes the smallest decisions, or whether it can also belong to the person who designs the larger musical system in which those decisions acquire their meaning.

So what?

Suppose the Copyright Office gets this wrong.

Then an extraordinary thing begins to happen. The better our creative tools become, the less authorship their users may be allowed to claim. Give a composer a primitive algorithm that requires her to specify every probability table and transformation, and we comfortably call her an algorithmic composer. Give her a more intelligent instrument that understands “thin the texture toward the cadence” without requiring fifty lines of code, and suddenly some of the expression has migrated into the machine. Improve the instrument further and still more authorship disappears. Eventually the very thing technology has always done for artists, allowing them to express intention at a higher level without manually constructing every detail, becomes evidence that they did not make the work.

That is not a small technical adjustment to copyright. It would establish a perverse bargain for the next generation of creators: you may use powerful tools, or you may own what you make with them, but the more powerful the tools become, the harder it may be to have both.

And who benefits from that?

Not necessarily the composer. Certainly not the young musician who discovers an entirely new way of working and then learns that the law recognizes only pieces of what she made. Perhaps the work still has commercial value. Perhaps people still want to hear it. Perhaps a platform can distribute it, a company can monetize access to the tool that made it, and everyone can conduct business around it. The only oddly uncertain participant is the human being who actually spent weeks deciding what the piece should become.

That is where this stops being a debate about whether we should be generous to AI users. I am not asking the Copyright Office to pretend a machine is human. Quite the opposite. This is not about giving machines authorship. It is about not accidentally taking authorship away from humans.

And it will not stop with the nut-job composers. Music just happens to make the problem unusually easy to see because we already have centuries of delegated realization sitting in plain sight. But increasingly, people will create by specifying relationships, constraints, intentions and systems while machines fill in details. That is what more capable creative tools are for. We will not become less expressive because we can operate at a higher level of abstraction. Quite possibly the opposite.

So before we quietly adopt a rule in which the human contribution shrinks whenever the tool becomes better at realizing it, we should make very sure that this is actually what copyright means.

The Copyright Office says it wants the traditional conception of authorship.

Good.

Then the tradition has a warning for it.

A composer has never had to make every decision.

A composer has had to make the decisions that make the work this work.

Sometimes that means choosing the notes.

Sometimes it means choosing what will choose the notes.

And if our newest instruments make that second kind of composing vastly more powerful, the question is not whether the machine has become an author.

The question is whether we are going to stop recognizing the actual human one.

Copyright © 2026 by Paul Henry Smith

There is something deeply strange happening in AI music right now. Companies are spending enormous effort negotiating licenses so their systems can legally work with music that somebody still owns, while at the same time preventing users from working with music that nobody owns anymore. If a company has licensed a catalog, fine: the rights holder has agreed, the company has permission, and the music can enter the product.

But try asking some of these same systems to make a new performance of Beethoven, Bach, Schubert or another composer whose work has long since entered the public domain, and the machine suddenly becomes very concerned about ownership.

Existing work detected. Existing artist detected. You may not have the rights. Please obtain permission.

From whom, exactly? Beethoven has been dead for nearly two hundred years. Bach for considerably longer. More importantly, nobody owns the compositions anymore. There is no permission left to obtain. The law already removed the gate. AI companies are putting another one up.

At first this looks merely stupid, the sort of clumsy overblocking we have all learned to expect from automated safety systems. But the more I have encountered it, the less trivial it seems, because it reveals how thoroughly we have forgotten one half of the bargain copyright was designed to make.

We talk now as though copyright exists because creators naturally own what they make, rather like owning a house or a violin, and eventually, for some obscure administrative reason, that ownership “expires.” But copyright is not ordinary property with a strangely short fuse attached to it. Copyright is a temporary exclusive right created for a public purpose. Society gives authors and composers something extraordinarily valuable: for a limited period, they may decide who copies, publishes, distributes, adapts and commercially exploits their work.

Why give them that power? Because making art takes time, skill and money, and a society that allows everyone else to copy a work instantly while its creator cannot earn from it may get less art. Copyright reconciles culture with commerce. It gives creators room to feed themselves.

But the word temporary is not decorative. It is the other half of the mechanism.

The deal was never that a work should remain under private control forever. The deal was that private control would last long enough to encourage creation and reward the creator, and then it would end. When that happens, the work is released from private ownership and enters the public domain. Not because it has become worthless. Not because society has forgotten about it. Not because somebody failed to renew the paperwork. The ending is intentional. Copyright has completed its job. The creator received the period of exclusivity society promised, and now society receives what it was promised in return: the freedom to use the work without asking anyone.

That public side of the bargain is easy to underestimate because “public domain” sounds rather like a municipal storage facility for culture that has outlived its usefulness. But look at what is actually there. Beethoven. Bach. Mozart. Schubert. Shakespeare. Dickens. And now, in the United States, works from 1930 including George and Ira Gershwin’s “I Got Rhythm.” This is not artistic landfill. Some of the most astonishingly alive material in our culture is in the public domain. It is played every night in concert halls, taught all day in music schools, used in movies and television, rearranged, quoted, borrowed from, argued about, recorded again and again. Music, in particular, does not become dead merely because it becomes old. Some of it seems almost perversely unwilling to leave.

This matters especially in music because performing an existing work is not some peripheral reuse of it. It is the central activity. A score is not a recording waiting to happen only once. It is information that performers continually turn back into music. Beethoven’s Fifth has been performed thousands upon thousands of times because no performance settles the matter. One conductor brings out a line everyone else has buried. Another takes a transition at a tempo that changes the entire shape of the movement. One pianist discovers something in Bach that another pianist somehow missed. We do not regard the five-hundredth recording of a Mozart concerto as an infringement upon the first four hundred and ninety-nine. We listen to see whether this musician has something to say. The performances are not from the past.

Musical education rests on the same freedom. A pianist spends years playing music other people wrote. An orchestra spends most of its life taking compositions that already exist and making them exist again. Children learn Bach. Students wrestle with Beethoven. Professionals return to Schubert after forty years and discover that they no longer hear the piece the way they did at twenty. The continual reuse of musical works is not a tolerated exception to musical culture. It is musical culture.

There is a legal distinction here that most people understandably never have reason to think about, and it explains a great deal of the confusion. A musical composition and a recording of that composition are not the same copyrighted thing.

If an orchestra records Beethoven’s Fifth today, that recording can be protected as a new creative work. I cannot take it, duplicate it, release it under my name and announce that Beethoven is in the public domain. But the existence of that protected recording does not somehow put Beethoven’s composition back under copyright. I can make another performance. You can make another performance. Another orchestra can make another performance tomorrow. Someone may own this recording of Beethoven, while nobody owns Beethoven.

That distinction is so ordinary in musical life that musicians barely notice it. Yet many AI systems behave as though it does not exist. They encounter a familiar work and see danger. There are copyrighted recordings of Beethoven’s Fifth in the world, therefore Beethoven’s Fifth itself begins to look proprietary. Existing music becomes functionally synonymous with owned music. The system asks for permission, but there is nobody from whom permission may legally be required for the underlying public-domain composition. Suno’s legal language, for example, recognizes the difference between having rights and obtaining permission. Their product does not.

And this produces an almost comic inversion. I can rent a hall, hire eighty musicians, put Beethoven on every music stand, rehearse the symphony, record it and release the result. I can play Bach on a piano in my living room and upload the recording tonight. I can arrange Schubert for instruments he never heard. I can make a truly appalling version of Mozart for accordions and tubas and discover, to my regret, that copyright law offers no protection from bad taste. I can “own” those recordings and copyright them. All of this is allowed because the music belongs to the public domain.

But introduce a generative music system, a new kind of instrument capable of making or helping make performances, and suddenly the same activity may be forbidden precisely because the music already exists.

Why?

The likely answer is not philosophical. It may be financial.

AI companies have very good reasons to be frightened of copyright litigation. If their systems allow someone to reproduce or transform protected music unlawfully, the consequences may be expensive. So they build filters. Recognizable song? Block it. Existing artist? Block it. Suspicious source material? Block it. And if the filter cannot reliably distinguish a public-domain composition from the many copyrighted recordings made from it, the economically rational thing is obvious: block that too.

A mistaken permission might produce a lawsuit. A mistaken refusal produces an irritated musician.

That explains the behavior. It does not make the behavior ethical.

And this, I think, is where we have developed a peculiar moral blind spot around copyright. We have become so fixated on one half of the equation, thou shalt not take what belongs to another, that we barely notice the other half: thou shalt not block access to what belongs to everyone.

We have built enormous moral, legal and technological systems around the first principle. If you use somebody’s protected song without authorization, alarms go off. Platforms scan for it. Lawyers write letters. Systems detect it automatically. Entire industries exist to administer rights and permissions. This is understandable. Private rights mean something only if they can be enforced.

But when private rights end, that must mean something too.

Suppose I walk into your house and take your guitar. The wrong is easy to see. The guitar is yours, and I have taken something I am not entitled to take. Now suppose instead that I put a fence around a public park and decide that nobody may enter without my permission. The wrong runs in the opposite direction. I have not stolen private property. I have asserted private control over something that is supposed to be available to everyone.

We seem to understand the first kind of wrong instinctively and have almost no moral vocabulary for the second.

Perhaps that is because one is visible. Copyright infringement looks like an act. Someone copied the recording. Someone distributed the film. Someone reproduced the photograph. There is an owner, an object and an apparent trespass.

Public-domain denial looks like nothing. The button is greyed out. The request fails. The system says no. You move on, assuming there must be a legitimate reason because decades of copyright enforcement have trained us to treat “no” as morally serious.

But sometimes the “no” is itself the violation.

The public domain is not a charitable gift that technology companies may choose to recognize when convenient. It is the outcome the copyright system was deliberately designed to produce. Strictly speaking, you and I do not possess little fractional deeds to Beethoven, but we possess something much more useful: nobody has the exclusive right to stop us from using Beethoven. No publisher, estate, corporation or individual may reacquire the composition and say, from now on, you need my permission again.

That absence of exclusion is the public’s property, in the meaningful sense.

So when an AI system encounters Beethoven and effectively says, you cannot proceed without authorization from the owner, it is not merely making an amusing historical mistake. It is applying the logic of private ownership to something that has deliberately ceased to be privately owned. It is demanding the permission of an owner who does not exist and, more importantly, must not exist, because the disappearance of that exclusive owner is exactly what copyright expiration was meant to accomplish.

Imagine if this were done legislatively. Suppose Congress announced tomorrow that Beethoven, Bach and Mozart were being placed back under copyright because new technology had made their music economically valuable again. Or suppose some corporation quietly obtained the power to license all future performances of Schubert. People would immediately recognize what had happened. The public’s side of the copyright bargain had been confiscated. Works that generations had waited to inherit had been put back behind a fence.

AI companies are not literally doing that. The legal status of Beethoven remains unchanged, and they are entitled to decide what their products will do. I still have other ways to perform the music. But inside what may become one of the most powerful new environments for making music, they can create much the same practical condition through product policy instead of law.

The gate is put back up.

And public-domain music is not even the only place where the law is more permissive than the machine. American copyright law has long contained a provision that requires rights holders to grant a compulsory license for many already-released songs. Once the statutory conditions are met, you or I can make a new recording of a copyrighted song, pay the required royalties and distribute it even if the copyright owner would prefer that I did not. That is not an oversight. The law deliberately separates compensation from veto power. Yet the emerging AI licensing regime can put the veto back.

Warner’s agreement with Suno promises participating artists and songwriters control over whether their compositions may be used in new AI-generated music. So we have arrived at something stranger than overzealous copyright enforcement: private agreements can make a new musical technology more restrictive than copyright law itself. A songwriter may be unable to stop me from lawfully recording her song with musicians in a studio, while being able, through a platform agreement, to stop me from making the equivalent new performance with AI.

I discovered how blunt this distinction can become when I uploaded a recording of myself playing Beethoven on my piano at home. Suno rejected it because, it said, “This audio matches an existing recording.” But it was not a copy of somebody else’s recording. It was mine. What it matched was, apparently, the fact that other recordings of the same work exist.

Beethoven’s composition has been in the public domain for generations; my performance was newly made by me. There was no third-party recording right to protect and no composition owner to consult. Yet the gate came down anyway.

This is why I do not think “they are just being cautious” settles the ethical question. Caution for whom? The rights holder is protected against possible infringement. The AI company is protected against possible litigation. The public, whose rights are also part of the copyright system, loses the ability to use lawful cultural material in a new creative medium because recognizing that ability creates no comparable financial incentive for the company.

One half of the copyright bargain receives elaborate technological enforcement.

The other half gets a greyed-out button.

That is not a balanced ethical framework. It is moral selectivity produced by asymmetric risk.

And the stakes become larger as the tools become more powerful. If generative music were merely a novelty that produced funny little songs, perhaps none of this would matter much. But these systems are rapidly becoming instruments for musical exploration. A musician can ask questions with them that once required performers, rehearsal rooms, studios, engineers and money. What happens if this inner voice comes forward? What happens if this phrase stretches further? What happens if the balance changes here? What happens if I make twenty interpretations of these eight measures and discover that nineteen are terrible but the twentieth reveals something I had never heard?

That is what new instruments do. They make previously expensive questions cheap enough to ask.

And public-domain music ought to be among the richest possible material for asking them. It is legally available. It is culturally inexhaustible. It includes works musicians have spent centuries continuing to discover. If there is anywhere generative technology should be able to operate with relatively little ethical ambiguity about the underlying composition, surely this is it.

Instead, we risk creating the first generation of musical instruments that become less willing to play a piece after its copyright ends.

Think how strange that is. A violin does not ask whether Bach has been licensed. A piano does not need approval from the Beethoven estate. A recording studio does not flash EXISTING ARTIST DETECTED when musicians arrive carrying Schubert. Notation software does not erase Mozart because Sony happens to own a recording of Mozart somewhere.

AI companies have built music systems so dependent on permission that they do not know what to do with music that is free.

And perhaps that is the larger lesson hiding inside this small absurdity. The language of “intellectual property” has become so dominant that we increasingly imagine copyright only as a mechanism for holding culture back from unauthorized use. We forget that it is also a mechanism for eventually releasing culture from authorization altogether. Its purpose is not merely to create owners. It is to make ownership temporary.

The public domain is society collecting its side of the bargain.

We already paid for Beethoven.

We paid by honoring the temporary monopoly. Generations accepted that for a time the composer, publisher or rights holder could say no. They could control copying, publication, adaptation and commercial exploitation because society believed that this temporary power would encourage the making of culture. Eventually that power ended, and the work came to us.

Not as a museum object.

Not merely so libraries could preserve it.

It came to us to use.

To perform. To record. To rearrange. To transform. To teach. To misunderstand and rediscover. To make ridiculous things from it and profound things from it. To place it beside technologies and ideas its creator could never have imagined and see what happens.

That freedom was not an accidental remainder after the important part of copyright was over. It was part of the reason the important part was temporary in the first place.

Which is why I find the prevailing notion of “ethical AI” here so upside down. We have become exquisitely attentive to the possibility that AI might let someone use a work that still belongs to somebody else, while displaying almost no concern when the same technology prevents someone from using a work that belongs to everybody.

Those are mirror-image failures.

In one, a person ignores a legitimate right to exclude. In the other, a company imposes exclusion where no legitimate right to exclude exists.

We should be capable of caring about both. But while we worried that AI would outrun copyright law, parts of the music industry instead are using AI to make private, copyright-like agreements that are more restrictive than the law itself ever was.

Ethics cannot simply mean minimizing the legal exposure of the company operating the machine. If our moral imagination stops wherever corporate liability stops, we have confused prudence with principle. The copyright bargain was made between creators and the public. AI companies arrived centuries later. They do not get to honor only the half of the bargain that protects them from lawsuits and quietly discard the half that gives the culture back to everyone else.

Copyright is temporary by design. Eventually, the lock comes off.

Society waited years for these works to become ours. It is a strange notion of progress to now build machines that lock them away again.

Copyright © 2026 by Paul Henry Smith

You know, it’s a strange thing, to be in a room and hear a child deliver a verdict. There’s a certain satisfaction in it, for everyone involved. The child gets a little pat on the head for being so aware, and the adult gets to feel as though they’ve successfully installed a tiny, functioning conscience. It’s a tidy transaction.

A child says, “AI is bad for the environment.”

And you think, all right. That’s probably true, in the way that most things are true if you say them with enough conviction. It certainly sounds like the sort of thing a thoughtful person would say.

But before we all start congratulating each other on a job well done, perhaps we could ask a small, rather impolite question. “How bad?”

The conversation stalls. The adult doesn’t really want the child to ask “how bad.” The adult wants the child to know that it is bad. The point isn’t to investigate the claim; it’s to acquire the correct attitude toward it, the way one acquires a taste for a certain brand of bottled water.

“It uses water,” the child will explain, with the solemnity of someone testifying before a grand jury.

Yes, it does.

So do hospitals. So do the factories that make the semiconductors in our phones, the power plants that charge our electric cars, the swimming pools we dip our children into on a hot day, the almond orchards that give us our milk alternative, the lawns we argue about with our neighbors, the craft breweries that give us something to do on a Friday night, the car washes that make our sedans feel momentarily new again, the golf courses my father plays on, and the cattle whose methane is, let’s be honest, a much bigger problem.

Now, we haven’t disproved anything. It would be absurd to say that because hospitals use water, data centers should be allowed to build a moat. But we have discovered that “uses water” is not, in fact, the entire argument. Everything interesting uses something. The question is how much, where, for what purpose, with what alternatives, and compared to what. Those aren’t loopholes. They are the beginning of a thought.

Suppose I tell you that a data center uses a million gallons of water. A million is an excellent number for this sort of thing. It has all those zeroes, which make it look as though something has gone terribly, terribly wrong.

But over what period? A day? A year? The entire lifespan of the facility? Is that water withdrawn, like borrowing a cup of sugar from the neighbor, or consumed, like eating the last of the ice cream? Is it drinking water? Reclaimed wastewater, the stuff that’s already been through the system once? Is it water evaporated through cooling towers, or water used indirectly to generate the electricity in the first place? And is the facility in rainy Oregon or drought-stricken Arizona? And what, for heaven’s sake, was the useful work it performed while it was using all that water?

If we don’t ask these questions, the number isn’t information. It’s stage lighting for a play we’ve already decided we don’t like.

What if we compare it with something familiar? The United States Golf Association estimates that American golf courses, collectively, use about 1.5 billion gallons of water per day. For putting a little white ball into a hole in the ground. Now, data centers use a lot of water, but estimates for all of them in the U.S. put the number at less than half of what we use for golf. And yet, children are not taught to recoil from a photograph of a putting green as if it were a crime scene. No teacher asks them to examine whether their parents’ country club membership is morally compatible with the watershed.

Why not? Perhaps golf is worth the water. People enjoy it. It provides employment, exercise, a place for businessmen to wear those terrible pants. The matter is not simple. Exactly. So why is AI not afforded the same elementary complexity?

The point isn’t that golf is wicked and AI is innocent. The point is that a child is being taught to apply one method of judgment to the novel, socially disfavored thing and another method to the familiar, socially protected thing. That is not education. It is target selection.

And then there’s the other certainty. “AI has stolen the work of artists.” It has a nice, biblical ring to it. A clear victim and a clear sin.

What exactly happened? Well, training a generative model can involve making copies of copyrighted works. Some datasets have included material that was, let’s say, ethically acquired. Some systems can reproduce or closely imitate a particular artist’s work, which can cause real economic harm. These are not invented concerns. But the legal and factual picture is not captured by the sentence “AI is theft.”

The U.S. Copyright Office, in a report, did not declare all generative AI training to be unlawful. It concluded that the answer depends heavily on the purpose of the use, the source of the material, the kind of model, and the effect on markets. Some uses might be fair; others might not. The question is contested. That is not a satisfying slogan. It is, however, the truth.

Human culture depends on learning from prior work. Painters study paintings. Writers unconsciously retain rhythms and structures. This does not prove that machine training is identical to human learning. It plainly is not. Machines operate at a different scale and economic power. But “not identical” is not the same as “obviously theft.” That is the question children ought to be allowed to inspect. When does learning become copying? When does copying become infringement?

But the strangest step is this: suppose a company did use unlawfully obtained material. Does it follow that every person who later uses the model is stealing? We don’t ordinarily reason that way. A person may buy a product from a company that has violated labor law. That may create an ethical question, but we wouldn’t say the consumer personally committed wage theft. To treat every use of AI as one moral act is to refuse to see any difference between asking for a grammar explanation and generating counterfeit books in the style of a living novelist.

There’s another problem with teaching a fixed moral picture of AI. The picture keeps going out of date. A child taught in 2026 that AI is inherently a distant, water-consuming industrial process may reach adulthood in a world where much of it runs invisibly inside their phone, their laptop, their car. What happens to the moral lesson then? Does it update? Or does the child preserve the disgust while forgetting its reason?

AI will stop announcing itself. It will be used to remove background noise from a podcast, to suggest an alternative word in an email, to organize your vacation photos. At which point does the thing become “AI-contaminated”? The child taught to reject AI as a visible category may discover that the category has disappeared into the environment.

So what is the child actually being taught? Perhaps the parent believes the child is learning environmental responsibility. Perhaps the teacher believes the child is learning respect for artists. But the child is also learning a method. A claim is presented. The claim contains a morally vivid word: theft, stealing, cheating, fake. No comparison is offered. No mechanism is inspected. No uncertainty is admitted. The child repeats the claim and receives social approval.

The lesson that has actually been learned is that a moral feeling can substitute for an investigation. And that lesson will travel. The child may apply it next to a political movement, a religion, a nationality, a class of people. Someone will provide the approved noun. Dangerous. Corrupt. Contaminated. The child will know what to do. Feel first. Repeat second. Examine never.

An honest lesson might begin with a list of claims and questions. Claim: AI uses too much water. Questions: Which systems? Which region? Compared with what? Claim: AI steals art. Questions: Was protected work copied? What does the law say? What should the law say?

Now the child has something better than a conclusion. She has handles. She can pick up the problem. The purpose of evidence is not always to reverse a judgment. Sometimes it makes the judgment more precise. A person may investigate a particular data center and oppose it more strongly afterward. A student may study a training dataset and conclude that its construction was exploitative. Excellent. That refusal now belongs to the person. It was not installed in advance.

The danger is not that today’s children will dislike AI. The danger is that they will be trained to maintain a moral belief after its factual supports have changed.

That’s where it starts to look less like education and more like a kind of faith, where every new piece of evidence is absorbed as proof of the conspiracy. The model becomes more efficient? That’s propaganda. The data is licensed? Consent under capitalism is impossible.

At that point, the person is no longer judging a technology. The person is defending a moral self-portrait.

Perhaps we should stop asking whether AI is good or bad. We don’t ask whether chemistry is good. Chemistry gives us fertilizer and nerve agents. We ask which chemical, which dose, which process. We might ask similar questions of AI. Which system? Which model? Running where? Trained on what? For whose benefit? Replacing whose labor?

Perhaps a child who learns to ask those questions will sometimes oppose AI more fiercely than her parents do. But she will know what she is opposing. And perhaps she will sometimes use it with delight, not because she’s been recruited into technological enthusiasm, but because she has found something worth doing.

Those questions do not provide the comfort of a verdict. They do something better. They leave the mind switched on.

Copyright © 2026 by Paul Henry Smith

Imagine that you are extraordinarily good at sailing across the Atlantic Ocean. Not merely good at traveling, but actually skilled at the business of crossing oceans: you understand navigation, currents, weather, ships, stars, all the things that once separated the person standing on one continent from the person capable of reaching another. It has taken you twenty years to learn this, and naturally you are proud of it.

Then airplanes arrive.

Now I come along knowing nothing about navigation. I can’t identify a current. I would be of very little use on a sailing ship. Nevertheless, I sit in a chair for seven hours, eat a meal, fall asleep, and wake up on the other side of the Atlantic.

Something has plainly been lost here. An accomplishment that once required rare knowledge can now be achieved by an ignoramus. If your livelihood depended on possessing that knowledge, you may have a serious problem. But suppose you complained about something slightly different. Suppose you said, “Now anyone can cross the Atlantic. What makes me special?”

We would know immediately that something had gone wrong in the reasoning. The airplane has certainly changed the value of your skill, but it has not somehow diminished the purpose of going to Europe. We understand without having to think very hard about it that there were always two things involved: the difficulty of crossing the ocean and the reason somebody wanted to be on the other side. Technology attacked the first one and, by doing so, became useful to the second.

With music, for some reason, these two things are much harder to separate.

A person spends twenty years learning harmony, counterpoint, orchestration, recording, performance. Then a machine appears into which someone can type a few sentences and receive a rather beautiful piece of music. If the music is bad, there is no philosophical problem; we can laugh at the machine and go home. The interesting case is the one that makes people nervous, so we should use that one. Suppose the music is excellent. Suppose eventually almost anyone can make it.

Now what exactly has become cheap?

It is tempting to answer, “Music.” But that answer is too quick, because there are very ordinary cases in which music matters enormously even though the musical information is already cheap.

Take a birthday party. A father sings “Happy Birthday” to his daughter. He did not compose the song, so we can remove originality from the equation. He is not a particularly good singer either; in fact, let him be embarrassingly bad. There are professional singers who could improve every measurable feature of his performance. We could replace him with one of them and get better pitch, better timing, better tone, probably a much nicer recording.

But we would not have improved the birthday.

This is a useful little result because it tells us that at least some of the value we casually attribute to music is not located in the musical object itself. The same sound can mean something quite different depending on who is making it, who is hearing it, and what is happening between them.

Once you notice that, other examples become almost annoyingly easy to find. At a concert, twenty thousand people sing a chorus they did not write, usually less accurately than the person onstage. Four friends in a car shout along with a recording they have heard fifty times. A pianist plays a piece written by somebody who died two hundred years ago, and an audience that already knows the piece comes to hear it again. A person sends a song written by a stranger to someone they love and somehow means, “This is what I have been trying to tell you.”

These are not marginal cases. They are a considerable portion of what human beings actually do with music. And they create a problem for the simple idea that the value of music lies primarily in producing original musical information. Much of the time, we seem to be using musical information to do something else.

What?

Well, perhaps music is a way of making feelings perceptible. Or a way of coordinating attention. Or a device for turning a collection of people into a group for three minutes. Those descriptions are incomplete, but they point in a useful direction: the artifact may be carrying the experience rather than constituting the whole of it.

This distinction matters because AI is extraordinarily good at producing artifacts. If we quietly assume that producing the artifact was the entire purpose of the activity, then a machine that can produce one cheaply appears to have swallowed the activity whole. But the birthday example tells us that this cannot be quite right. We can improve the artifact and worsen the thing we cared about.

At this point there is an obvious objection, and it is a good one. Surely the making matters. A pianist who has spent fifteen years learning an instrument has done something that a person typing instructions into a generator has not. The struggle changes your ear, your hands, your attention; perhaps even your character. Treating the two processes as equivalent would be ridiculous.

Agreed. But notice what follows and what does not.

It follows that learning the piano can be valuable. It does not follow that everyone who wants music must therefore learn the piano.

We have no difficulty making this distinction elsewhere. Walking across Italy may teach you things that flying over it cannot. You will see more, suffer more, notice more, and probably become a somewhat different person by the time you arrive. None of this requires us to conclude that the airplane is fraudulent transportation. We simply recognize that “getting to Rome” and “having the experience of walking to Rome” are different purposes.

Perhaps learning to compose and expressing something through music are also different purposes that happened, until recently, to require many of the same steps.

That possibility becomes more interesting if you look at the history of musical instruments themselves. Consider a piano. You move one finger a few millimeters and an elaborate mechanical system causes a felt-covered hammer to strike a carefully tensioned string. Someone else designed the mechanism, built the hammer, calculated the string length, constructed the resonating body and tuned the instrument. The pianist is taking advantage of an enormous pile of other people’s ingenuity every time she presses middle C.

This does not trouble us. We don’t accuse her of outsourcing the vibration.

So merely saying that technology performs part of the work cannot be the principle we’re looking for. Technology has always performed part of the work. A synthesizer performs more of it. A sequencer performs another part. Recording technology lets one performer become several. Editing lets us move a note after it has been played. Sampling lets us make music with sounds produced by people who are not even in the room.

There may certainly be a meaningful boundary somewhere between all this and asking an AI to make the whole song. But if there is, we ought to be able to say what the boundary protects. “The machine did too much” is not yet an explanation; it merely restates our discomfort in different words.

So let’s try another experiment.

Imagine that no AI was involved at all. Instead, a music teacher discovers a revolutionary teaching method. With this method almost anybody can become an excellent composer in six months. Harmony that once took years to understand can be grasped in weeks. Orchestration becomes intuitive. Millions of people who previously lacked the ability to compose can now do it beautifully.

Would that be bad for music?

It might certainly be bad for some professional composers. An increase in supply can reduce the economic value of a skill whether the cause is education, immigration, software, machinery, or AI. We shouldn’t wave that away. But there is another part of the situation that is harder to condemn. A teacher has found a way to give millions of people an ability that previously belonged to relatively few.

Ordinarily we call that a magnificent teacher.

This is where I began to suspect that the AI argument contains two different anxieties that are easily mistaken for one another. One concerns livelihood: what happens to me if a skill for which people once paid me becomes abundant? The other concerns identity: what happens to me if an ability that distinguished me becomes abundant?

The first question is economic and can be brutally real. The second is stranger.

We can examine it with an even simpler case. Suppose I am valuable because I know how to do something that you don’t know how to do. Now I teach you.

What happened to my value?

If the rarity of my ability is the important part, then I have diminished it. A teacher who gives away everything she knows is steadily manufacturing competitors. The greatest teacher would be the most self-destructive of all, because she would take knowledge concentrated in one person and distribute it as widely as possible.

Yet this is almost the reverse of how good teachers experience teaching. A mathematician does not normally despair when a student finally understands the proof. A piano teacher can be thrilled when a pupil plays something she once couldn’t play. Parents spend an astonishing portion of their lives teaching children to do without them the very things the children once depended on them to do.

There is a clue here. We seem perfectly capable, in these situations, of distinguishing between having something rare and having something valuable.

Civilization depends on the distinction. Almost everything we teach was once difficult knowledge possessed by very few people. Arithmetic, reading, navigation, musical notation, photography, programming: again and again we take an accomplishment that once distinguished a small number of specialists and make it ordinary. We don’t usually say that literacy was ruined when everybody learned to read.

What makes creative ability feel different?

Perhaps part of the answer is that we attach identity to it. “I compose” can gradually become “I am a composer,” which is not merely a description of an activity. It tells me something about my place among other people. I am the one who can do this. There is pleasure in mastery, certainly, but there is also pleasure in distinction, and it would be odd to pretend otherwise.

The trouble begins only when distinction has to carry more weight than it can support.

Suppose my sense of importance depends on being able to do something most people cannot. Then every improvement in other people’s abilities creates a tiny problem for me. A talented newcomer is a problem. A brilliant student is a problem. A better teaching method is a problem. A tool that lowers the barrier is a problem. The world becomes a peculiar zero-sum game in which spreading capability threatens the people who already possess it.

That seems like a very unstable arrangement for human worth. More importantly, it doesn’t describe some of the cases where human value is clearest.

Go back to the birthday party. The father is still singing badly. Why does he matter there?

It cannot be because he is unique in the sense of possessing a scarce capability. There are billions of people capable of singing that song. It cannot be because his performance is the best. We deliberately made it bad. It cannot even be because he created the thing.

He matters because of a relationship.

This is a different sort of specialness from the one threatened by abundance. If your oldest friend makes another friend, your friendship is not thereby diluted in the way an increased supply of diamonds can reduce the price of diamonds. If a million people become excellent composers tomorrow, your daughter does not value you one-millionth as much. These things don’t obey scarcity economics because their value depends on particular histories between particular people.

That suggests a distinction I find more useful than the familiar argument over whether AI art is “real art.” There is being special because few people can do what you do, and there is mattering to particular people because you are you. They can feel similar from the inside because both make us feel important, but they respond completely differently when capability becomes abundant.

And now the anxiety around AI starts to look different.

For most of history, the distance between imagining something and expressing it was enormous. You might hear a symphony in your mind, but to let anyone else hear it you needed years of musical education, knowledge of notation, access to instruments, performers, rehearsal time, perhaps a patron or an institution. A person could have something vivid to express and simply lack the machinery for getting it out.

Naturally, we admired the people who could cross that distance. How could we not? It was difficult.

But difficulty and value traveled together for so long that perhaps we stopped noticing they were separate variables.

You can see the confusion by imagining the opposite world. Suppose every person were born able to sing beautifully. Would singing to your child become meaningless? Suppose everyone could draw as naturally as they speak. Would a drawing made by your friend cease to be interesting? Suppose composing a song were as ordinary as composing a text message. Would people therefore have nothing worth saying with songs?

I don’t know the answers, and that is precisely why these seem to me more interesting questions than whether somebody “really made” an AI-assisted song. We are approaching a situation in which some forms of expression may become more detached from the years of technical training that historically preceded them. We have very little experience with that. We should expect some of our intuitions to be unreliable because they were formed in a world where the two were inseparable.

There is also a danger in becoming too enthusiastic about this. Removing the technical barrier may reveal that the barrier was doing useful things. Difficulty can force attention. Years of practice can produce judgment that cannot be obtained by requesting an output. A person who has learned harmony may hear possibilities unavailable to someone who hasn’t, even if both have access to the same generator. When production becomes effortless, we may discover that choosing well becomes harder, not easier. If you can have a thousand plausible songs before lunch, deciding which one deserves another person’s attention may become the scarce skill.

That would be an amusing reversal. We worry that AI will make creativity effortless, and then discover that it has merely moved the effort somewhere else.

Or perhaps it moves the value somewhere else.

Our world already contains a billion songs. It does not obviously need another song, no matter how that song is created. But people still need ways to say, “I miss you,” “This reminded me of you,” “Listen to this ridiculous thing I made,” or “I couldn’t explain what I meant, so I made you this.” The abundance of artifacts could make the relationships around them more important rather than less, just as the abundance of photographs did not eliminate our interest in photographs of people we love.

This is why I keep coming back to the airplane.

The interesting thing about an airplane is not that it makes everyone a navigator. It doesn’t. The passenger and the navigator remain different, and there may be excellent reasons to admire the navigator’s knowledge even in a world full of airplanes.

The airplane does something else: it separates the ability to cross an ocean from the ability to navigate one.

Generative AI may be beginning to separate things that we have also treated as one thing: technical mastery, artistic production, expression, originality, livelihood, status, and human connection. If so, arguing about whether AI “destroys creativity” may be premature. We haven’t even finished discovering which parts of creativity the machine is actually touching.

So when someone asks, “If anyone can make a great song, what makes me special?”, I don’t think the question needs a reassuring answer. Perhaps it needs to be taken apart.

What did learning to make music give you that remains valuable even if nobody needs you to manufacture the notes? What part of the pleasure came from mastery, and what part came from being unusual? If the ability to produce music becomes common, which of those things actually disappears? What new things become possible for people who previously had something to express but lacked the technique to express it? And if another person’s new capability really does make us feel smaller, what exactly have they taken from us?

Those questions don’t tell us whether generative AI will be good for music. They don’t tell us whether musicians will prosper, whether audiences will drown in synthetic rubbish, whether human performance will become more precious, or whether some forms of craft will wither when they are no longer necessary.

They do, however, give us a cleaner experiment to watch.

For a very long time, wanting to go somewhere and knowing how to get there were tangled together. Then we invented airplanes.

We are about to find out whether something similar happens when having something to express and possessing the technical ability to express it are no longer quite the same thing.

The interesting part may not be how many new songs we get. It may be what people turn out to do once more of them can cross the ocean.

Copyright © 2026 by Paul Henry Smith

You can hear Paul’s “made with AI” orchestral scherzos here: Carnival of Scherzos

Everybody knows the first temptation of a new tool: you try to make it do the biggest thing you have ever seen.

Give a child a hammer and he does not tap gently on a small nail. He wants to build a house, or maybe knock one down. Give a young musician an orchestra in his imagination and he does not say, “Let me write eight bars for a dance.” No, he says, “I will write a symphony.” Or an opera. There will be storms, kings, betrayals, death, moonlight, a chorus, and a great crash of brass at the end. It is thrilling. It is also usually no good.

And why is it no good? Not because the young composer lacks ambition. That is the easy explanation, but it misses the point. The real trouble is that he does not yet know what the smallest musical thing is. He can imagine a symphony, but he cannot write a minuet. He can imagine grandeur, but he cannot make one phrase lean into the next. He cannot make four bars ask a question and four bars answer it. He cannot make a tiny structure carry feeling.

That is a more serious failure than bad taste. Bad taste can be improved. But if you do not know what the smallest unit is, you do not yet know what medium you are working in.

This is where AI filmmaking is now.

People get hold of these extraordinary tools, and naturally they try to make the thing they have been trained to admire. They make Hollywood blockbusters. They make kung fu epics. They make trailers for impossible films: warriors flying through temple smoke, cities collapsing under purple skies, chrome robots kneeling in the rain, assassins turning in slow motion while sparks drift past their faces. And one should not sneer at this, because the excitement is real. Something that once required a studio, a stunt team, a render farm, and a ridiculous amount of money can now be approached by one person at a desk. That is astonishing.

But after the first astonishment, a small question appears. It is a dangerous little question because it ruins many impressive things. The question is: what exactly has been communicated?

Not displayed. Display is easy now. The machine can display almost anything. It can display a dragon, a palace, a burning moon, a thousand soldiers, a god made of glass, a child standing at the end of the universe. But communication is different. Communication means that something has passed from the film into the viewer with shape, pressure, and direction. An emotion has not merely been named or decorated. It has been made to occur.

And here is the hard part: the smallest unit of emotional communication in film is not the spectacular image. It is a knowing point. A moment in which the viewer knows something in a particular way.

A face changes, and we know the person has understood. A hand stops before touching a door, and we know the person is afraid of what is on the other side. A child lies, and the mother smiles, and we know she knows. A man looks at an empty chair, and we know who is missing. These are not big moments in the ordinary production sense. They do not require armies or explosions. But they are the atoms of cinema, because they create structured emotional knowledge in the viewer.

That phrase may sound too abstract, so let us make it plain. Emotion in a story is not just a mood. It is not fog, music, tears, shadows, or beautiful sadness poured over an image. Emotion is highly structured knowing. We feel grief because we know what has been lost. We feel suspense because we know something might happen and someone else does not. We feel tenderness because we know a character is trying not to show need. We feel dread because we know the ordinary room is no longer ordinary. The feeling comes from the arrangement of knowledge.

This is why so much AI film feels strangely weightless. It has emotion-colored surfaces, but not emotion-bearing structures. The image says “epic.” The lighting says “tragic.” The costume says “mythic.” The camera movement says “important.” But the viewer has not been given the precise knowledge required to feel anything definite. We are being shown the costume of emotion without the mechanism of emotion.

A kung fu film is not bodies flying through the air. That is the postcard version. A real kung fu scene is an argument in motion. We know who is stronger, who is faster, who is proud, who is afraid, who is hiding an injury, who refuses to retreat, who has discovered the opponent’s weakness. Every kick means something because it changes what we know. Without that structure, a fight is just acrobatics. With that structure, a single blocked punch can be thrilling.

A Hollywood blockbuster is not a large object exploding. It is a machine for distributing knowledge through time. The bomb under the table matters because we know it is there and the characters do not. The hero’s return matters because we know what it cost him to leave. The monster matters because we know what it represents to the child, the town, the family, or the coward inside the hero. Scale is not the thing. Time is the thing. Knowledge moving through time is the thing.

This is the part AI filmmakers must learn, and it is more radical than saying they need better judgment. Of course they need judgment. Everyone needs judgment. But judgment comes after one understands what one is judging. The deeper missing ingredient is a grasp of the elementary particle of the medium: the smallest moment that carries emotional knowledge from maker to viewer.

That is why AI film needs its Nickelodeon era.

Not because small films are morally superior to big films. That would be silly. Big films can be magnificent. But small films reveal whether the filmmaker understands the basic unit. Five minutes. One room. One desire. One reversal. One visual idea carried cleanly from beginning to end. A woman receives a message she cannot answer. A boy hides a broken toy from his father. A robot learns that a lie can be kind. A monster practices smiling in a mirror before meeting a child. These are tiny premises, but they force the essential question: what does the viewer know now that they did not know ten seconds ago, and how does that knowledge make them feel?

This is the minuet. This is the exercise that cannot be faked. In a five-minute film, you cannot hide behind lore, spectacle, world-building, or the promise that “this would be amazing as a feature.” The little film says: show me the emotional transaction. Show me the moment before and the moment after. Show me the point at which knowledge changes. If you can do that, then you are making cinema. If you cannot, then the dragon will not save you.

The old Nickelodeon films were simple not because their makers were simple-minded, but because the medium itself was being discovered one small mechanism at a time. A chase teaches one thing. A close-up teaches another. A cut teaches another. A gag teaches timing. A glance teaches inner life. The grammar of cinema was not born in full imperial costume. It was found by making little things that worked.

AI filmmaking has to go through the same humiliation, and the same joy. It has to stop dressing itself in the clothes of mature cinema before it has learned how cinema thinks. It has to stop asking, “Can I make something that looks like a blockbuster?” and start asking, “Can I make one image that changes meaning over time?” Can the same empty chair be funny at the beginning, worrying in the middle, and heartbreaking at the end? Can one doorway become a threat, then a hope, then a decision? Can a cup of tea tell us that someone was loved?

That sounds modest, but it is not modest at all. It is the whole art in miniature. Because once you can make one precise unit of emotional knowing, you can connect it to another, and then another. Now you have rhythm. Now you have story. Now the viewer is not merely looking at images, but traveling through a chain of recognitions. That chain is cinema.

The future of AI film will not be won by the person who can ask the machine for the most elaborate spectacle. That power will become common. The interesting artist will be the one who can use this absurd abundance to make something exact. Not a thousand warriors in the rain, but one warrior realizing he does not want to draw the sword. Not the exploding city, but the child who understands, from a flicker in her mother’s face, that home is already gone.

That is the beginning. Not bigger images, but smaller truths arranged in time. Not the imitation of Hollywood, but the rediscovery of the basic cinematic atom: a moment in which the viewer comes to know something, and because they know it, they feel.

That is what AI film is missing. Not the power to imagine. It has that in excess. It is missing the discipline of emotional structure, the knowledge of the tiny unit, the minuet before the symphony. And until it learns that, it will keep making magnificent empty rooms, waiting for a story to walk in.

The same mistake appears the moment people begin making music with AI.

Give someone a machine that can produce an orchestra, a choir, a trap beat, a film score, a jazz trio, a children’s chorus, and a Bulgarian folk ensemble before lunch, and what do they do first? Naturally, they try to make the biggest thing. They make the anthem. They make the trailer cue. They make the fake Bond theme, the fake Disney ballad, the fake Hans Zimmer mountain of brass, the fake K-pop chorus with all the furniture in the right place. And again, one should be a little kind about it. It is exciting. The machine opens a door that used to be locked by years of training, expensive rooms, hired musicians, and a great deal of tedious technical competence.

But the question is not whether the machine can make music-shaped sound. It can. The question is whether the music communicates.

And here we run into the same little trap as before. People think music communicates by sounding like music they already know. So if it sounds like a blockbuster score, it must be dramatic. If it sounds like a pop song, it must be catchy. If it sounds like a string quartet, it must be refined. If it sounds like a sad piano ballad, it must be moving. But that is like thinking a person is honest because they are wearing a suit. The costume may be appropriate, but the character is still unproven.

The smallest unit of music is not the genre. It is not the arrangement. It is not the drop, the chorus, the beat, the pad, the vocal texture, or the enormous cinematic drum that announces, with great confidence, that something important is happening. The smallest unit of emotional communication in music is a felt relation through time. One sound creates an expectation, and the next sound answers it, delays it, violates it, or deepens it.

That sounds technical, but it is the simplest thing in the world. Hum two notes. The first one leans somewhere. The second either gives it what it wants or refuses. That is already drama. A rhythm starts walking, and then it stumbles. A chord makes the room feel settled, and then one note inside it becomes restless. A melody climbs higher than it did before, and because we remember the earlier version, the new height means something. Music is not just beautiful sound. Music is organized wanting.

This is what much AI music does not yet understand, because many people using it do not yet ask it to understand. They ask for style, not necessity. They ask for “an emotional orchestral track,” “a nostalgic synthwave song,” “a dark cinematic cue,” “a catchy upbeat pop chorus.” The machine dutifully supplies the surface indicators. Minor chords. Reverb. Pulsing bass. Breathless vocals. Swelling strings. A riser. A beat drop. A melancholy piano figure that sounds as if it has been looking out a rainy window since 2011.

And it works for a few seconds. The ear recognizes the costume. “Ah,” it says, “this is sadness.” “Ah, this is triumph.” “Ah, this is danger.” But recognition is not the same as feeling. Feeling requires motion. It requires that something be set up, remembered, transformed, and made meaningful by its return. Without that, the music becomes a wallpaper of emotional signs. It tells you the room is haunted, but nothing in the room ever moves.

A good melody is not merely a pleasant string of notes. It is a little story about pressure. It begins somewhere, reaches, hesitates, repeats itself with a difference, finds or fails to find rest. Even a nursery rhyme knows this. Even a folk tune knows this. The tune is small enough that you can hold it in your head, and because you can hold it, the changes matter. If the melody comes back altered, you notice. If the rhythm arrives early, you feel the tug. If the final note lands where the whole phrase seemed to be going, you feel the pleasure of arrival, not because someone labeled it “satisfying,” but because your brain has been waiting.

This is also the minuet problem. The beginner does not want to write eight good bars. Eight good bars are merciless. They expose everything. If the first two bars do not imply the next two, the phrase dies. If the cadence is clumsy, the whole thing sits down in the mud. If the tune cannot be remembered, it cannot return with meaning. So the beginner reaches for scale. More instruments. More texture. More sections. More “production.” A vast glowing cloud of music-like music.

AI makes this temptation much worse, because production is now cheap. You can get the gloss of completion before you have earned the logic of a phrase. The track arrives mastered, arranged, and dressed for the stage. It has drums that hit, vocals that shimmer, bass that knows how to occupy the bottom of the mix, and little ear-candy sounds sprinkled around like sugar. But ask the impolite question: what is the musical idea? What is the one thing the piece knows how to do? What does it teach the listener to expect, and how does it reward or disturb that expectation?

Often the answer is: nothing very definite. It proceeds. It resembles. It fills time.

This is why AI music needs its Nickelodeon era too, or perhaps its minuet era, since music already gave us that better metaphor. Not three-minute songs imitating the most familiar structures of the streaming economy. Not fake albums. Not fake symphonies. Not endless “vibes.” Thirty seconds. Eight bars. One motif. One rhythmic trick. One harmonic turn. One emotional idea so small that it cannot hide behind production.

Make a tune that sounds brave the first time and foolish the second time. Make a two-chord loop that becomes heartbreaking because one note refuses to move. Make a drum pattern that teaches the body to expect a hit, then leaves a hole exactly where the hit should be. Make a melody that cannot finish its sentence until the very end. Make one sound enter as a joke and return as a memory. These are not small assignments because they are easy. They are small because they are exact.

The point is not to reject big music. Nobody who loves music could believe such a thing. Symphonies are wonderful. Operas are wonderful. Pop spectacles are wonderful. A great film score can lift an image into myth. A great dance track can make a crowd think with one nervous system. But the bigness is not where the music lives. The bigness is an accumulation of small intelligences. A symphony is not a large number of instruments. It is a chain of consequences. An opera is not singing plus tragedy plus costumes. It is a machine for turning desire into sound over time.

So the AI musician has to learn to ask smaller and more dangerous questions. Can this phrase be sung back after one hearing? Does this chord change alter the emotional temperature, or does it merely move to another approved chord? Does the rhythm create a body expectation? Does the return of the chorus mean more because of what happened in the verse? Is the silence placed where silence can do something? If the answer is no, then the track may still be pleasant, but it is not yet thinking musically.

The strangest thing is that AI may become most useful not when it gives us finished songs, but when it lets us experiment with these tiny units quickly. What happens if the melody rises instead of falls? What happens if the bass refuses the chord? What happens if the second phrase answers the first too soon, too late, or not at all? There is enormous creative power there, but only if the maker knows what is being tested. Otherwise the machine becomes a slot machine of styles. Pull the lever, get a vibe. Pull it again, get another vibe. This can be amusing, but it is not composition.

Composition begins when a sound makes a promise.

That is the simplest handle. A note promises. A rhythm promises. A chord promises. A texture promises. Not in words, of course, but in the nervous system. The listener begins to know what might happen next. Then the music plays with that knowledge. It confirms it, teases it, postpones it, contradicts it, or fulfills it in a way that feels surprising and inevitable at the same time. That is emotion in music: structured expectation becoming structured feeling.

So AI music is not missing the ability to sound impressive. It has that already, and soon it will have it in embarrassing abundance. It is not even mainly missing taste, though taste will always matter. It is missing, at the level of practice, a shared discipline around the smallest musical unit of communication: the phrase that wants, the rhythm that teaches, the return that remembers, the silence that changes the meaning of the sound around it.

Before the AI symphony, write the minuet. Before the fake album, write eight bars that can stand in the open air. Before the cinematic climax, write one musical question and one answer. Make the little thing work so clearly that no amount of production could improve its spine.

Then, and only then, let the orchestra enter.

Here we are: the beginner with AI tries to make the finished-looking thing: the poster, the concept art, the novel chapter, the proof, the product pitch, the polished essay, the “style.” But every medium has a smallest working unit, and if you do not understand that unit, the large thing becomes a costume.

In visual art, the unit is not “a beautiful image.” It is a relation: this shape against that shape, this edge against that softness, this patch of light telling the eye where to go. A picture communicates when it organizes seeing. The bad AI image often has plenty of detail, but no visual sentence. The eye wanders through ornamental fog.

In writing, the unit is not “a paragraph in the style of.” It is a thought moving cleanly from one mind to another. A sentence makes a little promise: here is something worth following. The next sentence must pay it off or complicate it. AI writing often fails not because it lacks vocabulary, but because it lacks necessity. The sentences resemble prose, but they do not have to be there.

And math is the most revealing case, because math has almost no patience for costume. The unit is not the grand theorem. It is the step that preserves truth. One implication. One definition doing real work. One example that shows why the claim might be true. One obstruction that shows why it might fail. A beautiful proof is not a cathedral made of symbols; it is a path where each stone bears weight.

So yes, the same rule holds: before the masterpiece, find the atom.

The danger of AI is that it can imitate the large form before we understand the small one. That is why it gives us instant symphonies, instant paintings, instant essays, instant proofs, and instant movies that feel oddly hollow. It has skipped the apprentice’s humiliating education in smallness.

But smallness is not a limitation. It is where the medium becomes real. A five-minute film, an eight-bar melody, a two-value drawing, a plain sentence, a one-line proof — these are not exercises beneath art. They are where art learns to stand up.

And yet this is why the whole thing is more hopeful than it first appears.

At the moment, much of AI art feels childish because it is childish. I do not mean that as an insult. Childhood is not a moral failure. Childhood is what happens when a new power arrives before the habits that discipline it. A child who has just learned to run does not begin by studying balance, gait, and long-distance pacing. He runs straight into the furniture. He runs because running is miraculous. Only later does he learn that movement is not the same as grace.

That is where we are. The tools have learned to run, and the people using them are running too. They are making fake blockbusters, fake symphonies, fake oil paintings, fake novels, fake proofs, fake everything. It is easy to mock this, because the results are often ridiculous. But ridicule is too cheap a response. The more interesting thing is to ask: what stage of development are we watching?

We are not watching the end of art. We are watching the baby pictures.

That changes the feeling of the whole matter. A baby’s first words are not literature, but they are not meaningless either. They are the beginning of a creature discovering that sounds can reach another mind. A child’s drawing of a person, with the arms coming out of the head and the sun smiling in the corner, is not bad anatomy. It is early communication. The child has count d the fact that marks can stand for things. That is enormous. The refinement comes later.

So perhaps AI culture is now in its arms-coming-out-of-the-head period. The images are full of fingers, fog, jewels, eyes, cities, impossible fabrics, tragic lighting, and solemn people looking at nothing in particular. The songs have all the instruments but no internal necessity. The films have the weather of emotion but not yet the event of emotion. The writing has the posture of thought before it has the pressure of thought. It is all a little embarrassing. But embarrassment is often what the beginning of learning looks like when seen from the outside.

The hopeful part is that this stage cannot last forever.

Why not? Because people get bored. That is the saving force. At first, the mere fact that the machine can make a dragon is enough. Then dragons become common. At first, the fake trailer is astonishing. Then everyone has made one. At first, the synthetic song that sounds almost like a hit feels magical. Then the ear begins to notice that it does not remember the tune. Abundance cheapens spectacle, and that is not a disaster. It is the condition that makes art necessary again.

When everything can be made large, largeness stops being impressive. Then the old questions return with more force than before. What is the idea? What is the turn? What is the smallest thing here that actually works? What does the viewer know? What does the listener expect? What does the reader understand now that they did not understand before? What step in the proof made the impossible become obvious?

Those questions are coming. They have to come. The babies will not remain babies.

And when these babies become teenagers, something fascinating will happen. They will still have the wildness of the new tool, but they will begin to feel shame. Not shame in the destructive sense, but the useful artistic shame that says, “Wait, this is too easy. This is a trick. This looks impressive, but it does not land.” That little discomfort is precious. It is the beginning of standards.

The teenager wants to be taken seriously. So the teenager studies. The teenager imitates, yes, but with more suspicion. He notices that the kung fu scene is not exciting because the camera spins, but because the fighter is learning something with his body. She notices that the sad song is not sad because the piano has reverb, but because the melody cannot quite return home. They notice that a good paragraph does not merely sound intelligent; it performs a clean act of understanding. They notice that a proof is not a string of official-looking symbols, but a sequence of necessities.

That is the moment when the medium begins to grow up.

The great surprise may be that AI does not destroy apprenticeship. It may make apprenticeship unavoidable in a new way. Because when production becomes cheap, the excuses disappear. You cannot say, “I failed because I did not have the budget.” You cannot say, “I failed because I did not have the orchestra.” You cannot say, “I failed because I could not render the world in my imagination.” The machine can give you all that, or something close enough to reveal the deeper problem. Now the failure is exposed at the level of conception. That sounds harsh, but it is also wonderful. It means we can finally see what needs to be learned.

The next generation of AI artists may therefore become connoisseurs of tiny causes. They will learn that one glance can carry more cinema than a collapsing planet. They will learn that one unresolved note can do more than a hundred tracks of strings. They will learn that one clean sentence can beat a page of perfume. They will learn that one honest mathematical step is more beautiful than a cathedral of notation. They will not reject scale. They will earn it.

And then the really interesting work begins. Because once the small unit is understood, AI’s abundance becomes powerful rather than distracting. Imagine a filmmaker who knows exactly what emotional knowledge must change in a scene, and can now explore twenty visual ways to make that change visible. Imagine a musician who understands the little promise inside a motif, and can now test a hundred orchestrations without losing the promise. Imagine a painter who knows the eye must travel from this hard edge to that soft shape, and can now invent worlds around that path without burying it. Imagine a writer who knows the thought, really knows it, and can now vary the rhythm, the metaphor, the angle of approach, while preserving the little engine of meaning.

That is not the death of craft. That is craft with a larger laboratory.

Of course there will still be oceans of junk. There always have been. The printing press did not produce only Shakespeare. Cheap cameras did not produce only cinema. Electric guitars did not produce only Hendrix. Every new medium multiplies rubbish before it multiplies masterpieces, because rubbish is what happens when power outruns understanding. But history suggests that understanding catches up. Not everywhere, not in everyone, not automatically in some noble march of progress, but enough. A few people become dissatisfied. A few people study the small thing. A few people ask better questions. Then the standard moves.

So the hopeful conclusion is not that AI will magically make everyone an artist. It will not. That was never how tools worked. The hopeful conclusion is stranger and better: AI may force the difference between making and communicating into the open. For a while, people will confuse the two. They will generate and generate and generate, covering the world in brilliant surfaces. But eventually the surfaces will cease to astonish. Then the old human hunger will remain: show me something that means.

And when that hunger returns, these baby forms will become adolescent forms, and then perhaps mature forms. The five-minute AI film will learn how to make us know. The thirty-second AI composition will learn how to make us expect. The AI image will learn how to direct the eye with purpose. The AI sentence will learn how to carry thought instead of merely wearing language. The AI proof will learn to move one necessity at a time.

That future is not guaranteed in every work, but it is hard to imagine it not appearing somewhere. Too many people will have the tools. Too many will get bored with spectacle. Too many will feel the itch that says, “There must be something more exact than this.” And that itch is where art has always begun.

So let the babies bang on the piano. Let them draw the arms out of the head. Let them make the dragons, the fake trailers, the synthetic operas, the neon samurai, the moody astronauts, the impossible palaces. That noise is not the final music. It is the nursery.

One day, sooner than we think, some of these babies will become teenagers. They will roll their eyes at the baby pictures. They will be embarrassed by the old dragons. They will want sharper tools, cleaner ideas, smaller units, stronger feelings. They will discover the minuet. And then, after enough minuets, someone will write the symphony.

Copyright © 2026 by Paul Henry Smith

A few nights ago, George Washington appeared in my living room and began inspecting my thermostat.

This was not the first thing I expected him to do. I had always imagined that if the Father of Our Country returned from the dead, he would appear on a battlefield, or in Congress, or perhaps in some quiet chamber of national reflection where portraits stare down at men pretending to be larger than themselves. Instead, he materialized next to a bookcase, wearing his uniform and an expression of stern concern, and asked whether the small glowing rectangle on my wall controlled the weather.

I told him that, in a limited domestic sense, it did.

He studied the thermostat with the grave attention one gives to artillery.

“And the fire,” he asked, “where is it kept?”

I explained central heating, which he received better than I expected. He had the manner of a man who had seen enough difficulties in life not to be rattled by a hidden furnace. Valley Forge, I suppose, prepares one for ductwork. Still, he seemed bothered by the absence of visible labor. In his world, warmth had to be coaxed from wood, flint, and human effort. In mine, it was summoned by pressing a plastic button with one’s thumb.

“That is convenient,” he said.

“It is.”

“And yet I detect no gratitude in your voice.”

This was awkward, partly because he was right and partly because he had been in the room for less than four minutes.

To recover my footing, I offered him coffee. He accepted, and then watched in silence as I placed a small capsule into a machine and made a noise that might have suggested to an eighteenth-century mind that a miniature prisoner was being steamed to death. When the coffee appeared, he examined it with suspicion, then tasted it, then brightened.

“This is excellent,” he said.

“It’s actually not considered very good coffee.”

He looked at the cup again, then at me.

“Your age must be a paradise of ingratitude.”

That is the difficulty with receiving visits from founders. One hopes to impress them with progress, but they keep noticing character.

Over the next hour I showed him various conveniences of modern life. The refrigerator delighted him, though he was suspicious of the freezer and asked whether it had diplomatic applications. The electric light impressed him, but not as much as the garbage disposal, which he listened to with the solemn pleasure of a man hearing a small battle go well. The television troubled him immediately. He did not like the number of people speaking at once, nor did he understand why so many of them appeared angry while seated indoors.

“Are they confined?” he asked.

“No, they’re being paid.”

“To quarrel?”

“Yes.”

He nodded slowly, as if some essential principle of the new republic had just slid several inches to the left.

I showed him my phone, which was a mistake. Oh, he accepted the idea that a small device could receive messages from great distances, but when I explained that it could also supply directions, display maps, record pictures, answer questions, summon strangers in automobiles, purchase shoes, identify birds, locate nearby Thai food, and allow citizens to broadcast their private disappointments to the world, he asked whether we had assigned it any limits.

I said that we had not, really.

“Then perhaps the device is the government,” he said.

I told him not to get ahead of himself.

Things were going reasonably well until I mentioned that the United States would soon celebrate its 250th birthday. At this, Washington straightened visibly. A quarter millennium meant something to him. I could see it in his face. He was not a vain man in the small sense, but he had gambled his life on an experiment that, at the time, looked about as practical as building a clock out of soup. To hear that it had lasted 250 years moved him more than the refrigerator, the thermostat, or even the garbage disposal, though I believe that last one remained in contention.

“Two hundred and fifty years,” he said, and for a moment his eyes moved past the room and into whatever interior country the dead are allowed to visit. “Then the republic endured.”

I hesitated.

There are questions that are easy to answer legally but difficult to answer socially.

“It did,” I said.

“And how will the occasion be marked?”

I had known the question was coming, yet still I found myself unprepared. One cannot simply throw a thing like this at George Washington. You have to approach it the way one approaches a horse with a musket wound, slowly and from the side.

“There will be events,” I said.

“Public ceremonies?”

“Yes.”

“Music?”

“Very likely.”

“Orations?”

“Unfortunately, almost certainly.”

He smiled at that, and I briefly loved him.

“And what else?”

I took a breath.

“There will also be a fight.”

Washington considered this. The word did not alarm him. His century had been well supplied with fights, and unlike ours, it generally knew what they were about. A fight, to him, implied some unsettled matter between parties. It suggested grievance, territory, insult, tax policy, sovereignty, or tea. He waited for the rest of the sentence, assuming that language, being civilized, would eventually provide the missing object.

“Who is fighting?” he asked.

“Professional fighters.”

“And for what cause?”

“They’re not fighting for a cause.”

“Then for whom?”

“Not really for anyone.”

“For land?”

“No.”

“For honor?”

“Not in the older sense.”

“For command?”

“No.”

“For settlement of a dispute?”

“No dispute.”

“For money?”

“Yes, but also for attention.”

He looked relieved to hear the word money, which at least belonged to recognizable human weakness, then confused again when attention was added to it. Money he understood. Attention seemed too vague to be a motive and too childish to be a public principle.

“So the fight is an entertainment?”

“That’s right.”

“And this entertainment is part of the anniversary?”

“Yes.”

“Of independence?”

“Yes.”

“At the president’s house?”

“Yes.”

He leaned back. “I see.”

He did not see.

I could tell because he had the expression of a man who has decided to postpone understanding out of courtesy.

I then made the tactical error of explaining that the fight would take place on the South Lawn of the White House beneath an enormous structure called The Claw.

At the mention of this name, Washington’s face changed. It was not fear exactly. It was the look of a man who had expected civilization to contain errors but not mascots from a seafood nightmare.

“The Claw,” he repeated.

“Yes.”

“Is this a military implement?”

“No.”

“A monument?”

“No.”

“A symbol?”

“Technically, everything is a symbol if it gets large enough.”

He ignored this, wisely.

“What does it do?”

“It holds lights and screens and speakers and various equipment.”

“For the fight?”

“Yes.”

“And it is called The Claw?”

“Yes.”

“By whom?”

“That’s harder to answer.”

“Did Congress name it?”

“No.”

“Did the people vote upon it?”

“No.”

“Did someone lose a wager?”

“Possibly, but not officially.”

I showed him a picture. He studied it for a long time. The image displayed the South Lawn with the immense apparatus rising over it, an octagonal cage beneath, the White House in the background looking less like the seat of executive authority than a tasteful historical prop rented for the afternoon. Washington did not speak for nearly a minute.

“At Mount Vernon,” he finally said, “we had a smokehouse.”

“That’s different.”

“It too had a purpose.”

This was unfair, but devastating.

I explained that the purpose of The Claw was to create atmosphere. Atmosphere is one of those words we use when the truth would sound too simple. We no longer decorate a room. We curate a vibe. We no longer illuminate a stage. We design an experience. We no longer construct a giant claw-shaped superstructure above the lawn where presidents have walked and children have hunted Easter eggs. We activate a national moment.

Washington asked, “what is being activated?”

“The anniversary.”

“Are anniversaries customarily inactive?”

“Not exactly.”

He asked why this one required mechanical assistance.

At this point I offered him more coffee.

He accepted, but I noticed he kept the photograph in his hand, as though by staring at it long enough he might discover the missing principle.

“What are these names upon the fighting enclosure?” he asked.

“Sponsors.”

He repeated the word. “Sponsors.”

“Companies that pay to associate their names with the event.”

“With the fight?”

“Yes.”

“With the anniversary?”

“Yes.”

“With the president’s house?”

“Also yes.”

He looked at the photograph again. I saw him reading the names. Tires, beer, wireless service, legal assistance, crypto. Nothing improves a birthday like tire branding. Nothing says two and a half centuries of self-government like a law firm name displayed near a cage where trained men prepare to strike one another for public enjoyment.

Washington pointed to the steps leading into the Octagon.

“And these words here?”

“A company name.”

“On the steps?”

“Yes.”

“Does the company own the steps?”

“No.”

“Did it build the steps?”

“Probably not.”

“Does it improve the steps?”

“No.”

“Then why is it there?”

“Because it paid to be there.”

“For the honor?”

“For the visibility.”

This word, visibility, troubled him more than sponsorship. It had a faintly moral sound, as though we were discussing widows or the poor, yet here it referred to a staircase having been rented by a financial enterprise. I tried to explain the modern attention economy, which is like explaining gout to a trout. The more I spoke, the more I heard myself sounding like the press secretary for a disease.

“In our time,” I said, “attention is extremely valuable.”

“Because it leads to judgment?”

“No.”

“To wisdom?”

“No.”

“To action?”

“Sometimes, but usually not.”

“Then why is it valuable?”

“Because it can be sold.”

He placed the photograph on the coffee table with care, as one sets down a plate containing something not yet identified by science.

It was then that I understood the true problem. I had been trying to explain the cage fight as an event, but it was not really an event. It was a diagram. Every piece of it labeled another piece of us. The lawn, the cage, the sponsors, the lights, The Claw, the birthday, the president, the fighters, the cameras, the viewers, the brands crawling over the whole enterprise like ants at a picnic. It was not merely something America was doing. It was something America had become legible enough to do without embarrassment.

Wanting to give Washington a broader context, I told him this was not unusual. We often add spectacle to things now. He asked what sorts of things, and I said music, which seemed safe for approximately one second.

He liked music. This pleased me, because it gave us a bridge back to civilization. He knew that music could accompany ceremony, labor, worship, dancing, mourning, marching, and drinking, which covers a large part of the human enterprise. He assumed, as any sane person might, that if our technology had improved so dramatically, our music must have gained new powers of expression. I wanted to tell him yes. In some cases I could have done so truthfully. There is wonderful music being made, of course. But that was not the subject at hand, and the dead do not visit one’s living room to hear evasions.

So I described the modern concert.

I told him about arenas filled with tens of thousands of people. I told him about amplification, not merely so that music could be heard, but so that it could press upon the body like thunder. I told him about lights programmed to hit the audience in perfect coordination with musical events, or with events that seemed to have been placed in the music for the convenience of the lights. I told him about smoke machines, fog banks, gigantic screens, moving platforms, wristbands that illuminate in synchronized colors, and flames that burst from the stage in carefully timed columns, as though the music had discovered oil beneath itself and begun drilling.

He listened with growing admiration.

“And the music,” he said, “must be powerful indeed to command such attendants.”

This was precisely the trap. He had assumed that the spectacle served the music in the way a frame serves a painting or a procession serves a coronation. He thought the attendants had gathered because the monarch was worth attending. I could see no way forward but through.

“Sometimes,” I said, “the music is powerful.”

“And other times?”

“Other times the attendants do a great deal of the work.”

“How so?”

“Well, if the music itself does not create enough excitement, excitement can be supplied.”

“Supplied by whom?”

“Technicians.”

He blinked.

“And fire?”

“Yes.”

“And lights?”

“Yes.”

“And smoke?”

“Yes.”

“And giant images?”

“Yes.”

He thought about this.

“So the music is ill?”

“No.”

“But it requires assistance?”

“Not exactly.”

“Does it travel?”

The question startled me because it went directly to the center. I had not used the word travel. Washington had found it on his own. Perhaps in his century, before recordings and streaming and playlists and background music in pharmacies, music was still understood as an event that occupied time by moving through it. A song began in one state and ended in another. A march went somewhere because men were marching. A dance went somewhere because bodies were changing place. A hymn went somewhere because it asked the soul to rise, even if only as far as the rafters.

“Not always,” I said.

He waited.

“A great deal of modern music establishes a beat and remains with it.”

“For dancing?”

“Sometimes.”

“Then the dancers travel?”

“Not really.”

“Then perhaps the feeling travels?”

“Sometimes.”

“And when it does not?”

I looked at him. He looked at me. I thought of the concert I had seen where the stage had erupted so frequently that a person arriving late might have assumed the band was being attacked by dragons. The beat had not moved. The harmony had barely moved. The melody, insofar as one existed, had moved in the manner of a moth against a window. Yet the air had been full of light and smoke and heat. The crowd had cheered each eruption with the grateful relief of people being reassured that something, somewhere, was happening.

“When it does not,” I said, “we have production values.”

Washington repeated the phrase with obvious distaste.

I described one moment in particular, when flames shot up during a drum hit and the audience responded as if Moses had returned with an encore. It had made me wonder whether the next development in music might be the introduction of live animals. A flaming rhinoceros, for example, catapulted across the stage, would almost certainly produce excitement. One could hardly deny it. The creature would have movement, force, danger, novelty, and a tragic grandeur not always supplied by the chorus. Properly sponsored, it might even save the industry.

Washington, who had maintained admirable composure through the thermostat, the coffee machine, the phone, the television, The Claw, and the sponsored staircase, now raised a hand.

“Why would the rhinoceros be aflame?”

“To create excitement.”

“Would the song not create it?”

“That is the question.”

“Would the beast survive?”

“Probably not.”

“Then the music must be very poor.”

This was harsh, but I could not honestly call it wrong in principle.

I tried to soften the matter by explaining that many artists are under enormous pressure to hold audience attention in a crowded media environment. People are distracted. They carry devices. They expect stimulation. They have been trained by platforms, feeds, advertisements, and entertainment systems to become restless almost instantly. A song that asks to be followed must compete with a thousand other things that ask only to be noticed. Washington listened, then asked whether the audience had become weaker or the songs had become lazier.

I told him this was a complicated question.

“That usually means both,” he said.

The man was becoming a problem.

At dinner, which I ordered through my phone because I was beginning to lose moral authority anyway, we watched the news. I had hoped for weather, perhaps a human-interest story about a rescued dog, anything that might let the evening rest. Instead, the screen presented the actual war. The tone changed, at least superficially. The anchors lowered their voices into that deep professional register reserved for events whose seriousness must be indicated without reducing the attractiveness of the broadcast. Maps appeared. Red arrows moved. Smoke rose from buildings. Experts joined from remote studios, each in a box, like minor saints in a digital altarpiece.

Washington leaned forward.

“Is this current?”

“Yes.”

“And these are real people?”

“Yes.”

“Is the nation at war?”

“Not exactly.”

He did not like that answer.

I explained, badly, the modern condition in which a country may be involved, implicated, aligned, opposed, supportive, strategic, concerned, prepared, monitoring, funding, advising, and deeply engaged without necessarily using the simple older word war in a way that would satisfy a farmer from Virginia. Washington had experience with war as mud, hunger, disease, supply shortages, letters, horses, fear, and decisions that could not be revised after breakfast. Our televised version appeared to him suspiciously organized.

The war broke for commercials.

This, I realized too late, was another trap.

An advertisement for a hamburger appeared first, enormous and lacquered, the bun shining with the erotic confidence of an empire at peace with itself. Then came a truck driving through a desert for reasons that were not immediately clear, followed by a pharmaceutical commercial in which attractive people with an unnamed condition smiled near water while a voice listed possible side effects with the serenity of a priest reciting genealogy. Then the war returned.

Washington had not moved.

“Was the war interrupted by food?”

“Yes.”

“And medicine?”

“Yes.”

“And a carriage?”

“A truck.”

“Does the war resume where it left off?”

“In a sense.”

He watched for another minute.

“Do the people watching understand that one thing is not the other?”

“I think so.”

“Are you certain?”

This was an excellent question, and like many excellent questions, not one I wished to answer.

The war, the concert, and the cage fight began to gather in the room as though they had been invited separately and discovered they were cousins. I could feel Washington sensing the connection before I had the courage to name it. In each case, something intense had been placed inside a machine that required constant stimulation. Music, unable or unwilling to move the listener by musical means, acquired flames. War, too serious to be entertainment but delivered through the apparatus of entertainment, acquired graphics, pacing, panels, interruptions, and the polite hygiene of distance. The cage fight, relieved of any purpose beyond spectacle, acquired the White House, sponsors, and The Claw. Each thing entered the same container. Each was translated into the same language.

Washington did not accuse us of decadence. That would have been easier to endure. He did something worse. He asked practical questions.

“If the music does not move, why not write better music?”

“We do.”

“Then why the flames?”

“Because people expect them.”

“If the war is grave, why interrupt it with hamburgers?”

“Because broadcasting requires revenue.”

“If the anniversary is solemn, why hold a fight?”

“Because solemnity does not rate as well.”

“If the house is symbolic, why cover its lawn with advertisements?”

“Because symbols attract attention.”

“And attention is sold.”

“Yes.”

He sat back.

“I begin to understand.”

This was a dangerous development. I had hoped he would remain confused. Confusion can be charming. Understanding, in a visitor from the eighteenth century, is fatal.

“What do you understand?” I asked.

“That your nation has become enormously skilled at summoning feeling without providing an object worthy of it.”

I wanted to object, but he had caught us cleanly. We have become expert at producing the conditions that used to accompany significance. Lights, music, scale, urgency, noise, conflict, graphics, flames, uniforms, brands, slogans, countdowns, breaking news, historic announcements, limited-time offers, exclusive access, national moments, emotional arcs, immersive experiences, premium visibility. We know how significance looks from across the street. We know its costume, its lighting, its entrance music. What we have less confidence in is the thing itself.

The concert says: if the music cannot carry you, the flames will.

The broadcast says: if the war cannot hold you, the graphics will.

The politician says: if the argument cannot persuade you, the rally will.

The anniversary says: if history cannot move you, the cage fight will.

And The Claw, rising above the White House like a monument to the American talent for missing the point at industrial scale, says: if the republic cannot inspire reverence, perhaps production design can.

By now it was late. Washington had removed his coat and was sitting in my chair with the exhausted dignity of a man who had been forced to review a quarterly business plan for Hell. I asked whether he wanted anything else. More coffee? Water? A blanket? He asked if I owned a map. I brought him an atlas, which he opened with visible relief. Maps belonged to his world. Rivers, mountains, borders, roads, ports. Here was information that still had the decency to remain itself.

He turned to the page showing the United States.

“So large,” he said.

“Yes.”

“So wealthy.”

“Yes.”

“So capable.”

“Yes.”

He traced the outline of the country with one finger.

“And yet it celebrates itself by arranging a fight beneath a claw.”

“That is one way to put it.”

He smiled, and for the first time all evening there was something mischievous in it, though I would not have used that word aloud in his presence, as he looked like a man who considered adjectives a privilege.

“Perhaps,” he said, “you are being too severe.”

This surprised me.

“You think so?”

“It may be a useful ceremony.”

“How?”

“In my day, we had ceremonies to reveal ideals. Yours may reveal appetites.”

I had to admit this was fair.

He continued, warming to the idea. “A republic should know what it worships. If it worships courage, let it show courage. If it worships liberty, let it show liberty. If it worships wisdom, let it show wisdom. And if it worships attention, then by all means let it build the largest claw it can afford and place beneath it two men who are paid to injure one another while merchants write their names upon the steps.”

There are moments when sarcasm reaches such altitude that it becomes indistinguishable from public policy.

Washington seemed pleased with his formulation. I was not. It had the disadvantage of being both ridiculous and exact.

The next morning he was gone. Ghosts never help clean up. His coffee cup remained on the table beside the newspaper photograph of The Claw. The thermostat glowed quietly on the wall. My phone contained sixteen notifications, none of them necessary. The television, when I turned it on, had resumed its labors. There was war, then weather, then a commercial, then a panel discussing whether the upcoming White House fight would be good for the president’s image, then a segment about summer travel delays, then more war. Everything was where I had left it, which is to say everywhere at once.

I looked again at the picture of the South Lawn. In daylight, the structure was almost funny. That may be its greatest protection. The truly grotesque often survives by being amusing enough to avoid immediate resistance. A giant claw over the White House sounds like satire until the riggers arrive. A cage fight for the nation’s birthday sounds like a joke until the sponsors are finalized. A culture that requires flames to supplement music and commercials to finance war coverage and combat to enliven patriotism sounds like an exaggeration until one notices that the exaggeration has already purchased naming rights.

Still, I found myself grateful to Washington, or to whatever guilty corner of my imagination had dressed itself as Washington and visited. He had not given me a lecture. He had merely tried, like a reasonable person, to understand the country according to its own explanations. That was the cruelest method available. Nothing exposes a civilization faster than explaining it politely to someone who still assumes words have meanings.

The cage fight has not happened yet. We still have that to look forward to. Soon the lights will come on. The screens will glow. The brands will shine. The crowd will gather. The cameras will find their angles. The announcers will locate the proper tone of reverent excitement. The cage will wait beneath The Claw on the lawn of the people’s house, and somewhere in the surrounding darkness, actual wars will continue, actual songs will fail to move, actual citizens will search for something large enough to feel like meaning.

Perhaps the event will be a success. By our standards, it almost certainly will be. It will be visible. It will be discussed. It will be clipped, shared, condemned, defended, monetized, memed, and folded into the next day’s programming. It will generate exactly what it was built to generate.

Washington, I think, would understand that now.

He might not approve, but approval is beside the point. Approval belongs to an older world in which citizens still asked whether a thing was worthy of them. We have improved upon that. We ask whether it can hold attention.

And if it cannot, there is always the flaming rhinoceros.

Copyright © 2026 by Paul Henry Smith

One of the odd pleasures of the past few years has been watching the guardians of culture discover, with theatrical horror, that culture can now be made by people who were not invited to the meeting.

The panic has many costumes. Sometimes it appears in legal robes. Sometimes it arrives as concern for artists. Sometimes it speaks in the language of safety, transparency, authenticity, or ethics. But beneath these costumes is the older and more reliable animal: institutional self-interest, startled out of its nap.

When generative AI first entered public life, the owners of large catalogs of intellectual property faced a genuine problem. Here was a new class of tools that could write, draw, compose, narrate, edit, imitate, remix, summarize, and accelerate creative work with unnerving speed. It did not make everyone a genius, despite the claims of its more excitable salesmen. It did something more disruptive. It made competence cheaper, experimentation faster, and certain old tollbooths easier to avoid.

No established industry smiles at that sort of thing. A publisher prefers publishing to remain difficult. A record label prefers music production to remain difficult. A studio prefers filmmaking to remain difficult. A guild prefers expertise to remain scarce enough to be policed. This is not a conspiracy. It does not require a secret room, a tray of pastries, and a man from legal whispering into a speakerphone. Institutions with similar interests tend to reach similar conclusions, especially when the alternative is watching their leverage leak away.

The first useful conclusion was that AI needed a stain.

The word chosen was theft.

It was an excellent word for the purpose. Theft does not invite curiosity. It ends conversation. Once a thing has been successfully called theft, the public is spared the tiresome burden of asking what was taken, from whom, under what law, with what injury, and by what analogy. The word carries its own gavel.

Whether AI training is legally theft remains unsettled. The courts have not issued the sweeping moral verdict that the rhetoric often implies. The law is still busy doing what law does: moving slowly through distinctions that slogans trample flat. But the cultural verdict did not wait for the legal one. Culture rarely waits for evidence when a satisfying accusation is available.

And so the idea spread. AI training was theft. AI tools were built on theft. AI art was theft-adjacent. AI writing carried the fumes of theft. Even when a particular output copied nothing and infringed no identifiable work, it could still be treated as contaminated by ancestry. The sin had moved upstream. It was no longer necessary to show that the thing produced was unlawful. It was enough to suggest that the machine had once looked upon forbidden material, and that every later creation bore the family curse.

That was the first achievement: not a ruling, but a mood.

The second achievement was more subtle. It divided creators.

Before AI, at least in theory, a work had to stand in the open and take its beating like everything else. A song could be good or bad. A painting could be alive or dead. A novel could reveal the world or merely rearrange furniture in the usual dreary rooms. The creator might be overrated, underpaid, fashionable, derivative, inspired, fraudulent, or dull, but the work itself remained the main exhibit.

AI changed the location of the trial. The artifact stepped aside and the process took the witness stand.

A person could make something original, useful, funny, beautiful, or strange. It could copy no protected work. It could delight an audience. Then, if AI involvement were discovered, the discussion would curdle. The question would no longer be, “Is it good?” It would become, “But how was it made?” And behind that question lurked another one, less polite and more consequential: “Does this person count?”

This is the part worth pausing over, because it is the hinge of the whole affair. The deepest argument about AI is not really about originality. It is about legitimacy.

Who gets to be called a creator? Who is merely a user? Who is experimenting, and who is cheating? Who is assisted, and who is disqualified? The modern world runs on tools so deeply embedded that they have become invisible, yet AI alone has been singled out as the tool that stains the hand. Nobody demands a moral warning label on autofocus. Nobody says a song loses its soul because the producer used pitch correction, digital editing, sample libraries, or software instruments. Nobody asks whether a novelist’s word processor has corrupted the sanctity of prose. But AI, because it arrived suddenly and threatened existing hierarchies, has been assigned a special metaphysical odor.

Then came the next word: slop.

This one was even better than theft because it required no argument at all. Slop is not a legal claim. It is a sneer in a paper hat. It allows the speaker to dismiss the work, the tool, the maker, and the audience in one economical motion.

The joke, of course, is that human beings were producing slop long before a transformer model ever completed a sentence. The world was not a cathedral of taste in 2021. The internet was already full of bad essays, dead-eyed stock photography, cynical pop songs, fake sincerity, recycled jokes, vacant lifestyle content, and videos assembled for the apparent purpose of making time feel insulted. If slop is the crime, humanity has been caught standing over the body with a ladle in its hand.

But the new rhetoric performed a tidy little trick. Bad human work remained merely bad work. Bad AI-assisted work became evidence against AI itself. And if AI-assisted work turned out to be impressive, the charge changed shape. Then it was soulless, uncanny, too smooth, too competent, too suspiciously free of the noble bruises by which we prefer to recognize human effort.

This made acquittal nearly impossible. If the work was poor, AI had debased creation. If the work was good, AI had counterfeited creation. The user could not win, because the point was not to judge the work. The point was to keep the user in a lower caste.

Now we have entered the administrative phase, which is where modern moral campaigns go once the shouting has been converted into policy. The words are disclosure, labeling, watermarking, detection, transparency, authenticity, and trust. These are fine words, buffed to a civic shine. No one wants to be against transparency. It sounds like sunlight, clean glass, and responsible adults in conference rooms.

But labels do more than inform. They sort.

A label creates a class of things. A class of things can be ranked differently, recommended differently, demonetized differently, searched differently, suppressed differently, appealed differently, and explained to advertisers differently. Once AI involvement becomes a detectable condition, the platform gains a new switch. It can say, “This may be allowed, but not promoted.” Or, “This may be visible, but not monetized.” Or, “This may remain online, but only after wearing the little yellow badge of synthetic suspicion.”

Here the interests of large rights holders and large platforms do not need to merge into conspiracy. They merely need to overlap. Music labels already have arrangements with platforms. Their catalogs are licensed, surfaced, tracked, monetized, and made available as raw material for users’ videos. YouTube did not become one of the largest music services in the world by ignoring rights holders. It became one by building machinery that could accommodate them, pay them, identify their works, and keep enough peace for the whole bazaar to remain open.

Once such machinery exists, it can be adapted. The platform has every incentive to classify, filter, and manage risk. The rights holder has every incentive to preserve the status of its own catalog and make life harder for unlicensed competitors, outsiders, and amateurs with suspiciously capable tools. Neither side needs to twirl a mustache. Each simply pulls the levers it already controls.

The trouble is that the levers are attached to a machine that is aging quickly.

AI is not remaining a separate toy called “AI.” It is sinking into the floorboards of production. It is becoming an option in the editor, the camera, the audio workstation, the design app, the writing tool, the translation layer, the color corrector, the compression system, the search box, the thumbnail generator, the captioning tool, the recommendation engine, and the feature nobody notices until it stops working. It will not always announce itself with a glowing button and a wizard icon. It will be buried in the workflow, doing small useful things by the thousand.

A musician may use AI to sketch a bass line and then record it with a human player. A filmmaker may use AI to find the strongest sixty seconds in an hour of footage. A podcaster may use AI to clean background noise, generate a transcript, translate the episode, suggest clips, and produce show notes. A designer may use AI to extend an image, test layouts, name layers, and create variants. A teenager may make a video that is “AI-assisted” in six different ways without ever thinking of it as an AI project.

What, exactly, is the label supposed to label?

At first, the distinction seems obvious. Some content is AI-generated, and some is not. Then the tools improve. Then the features spread. Then the category begins to blur at the edges. Then it blurs everywhere. The question stops being whether AI was used and becomes whether its use was meaningful, material, deceptive, incidental, aesthetic, administrative, invisible, or simply part of ordinary software. At that point, the old binary starts to look less like a principle and more like a temporary filing system maintained by anxious clerks.

There is another problem. Audiences are not always as pure in their preferences as moral campaigns require them to be. People often like what they like before they know how it was made. Tell them afterward that AI was involved, and you may create not enlightenment but dissonance. They enjoyed the song, laughed at the video, used the image, learned from the explanation, or shared the clip. Then the label arrives, wearing its little official shoes, and informs them that their pleasure may require adjustment.

That may work for a while. Shame is a powerful user-interface pattern. But it is a dangerous business to train audiences to distrust things they plainly enjoy, especially when the distrusted method is becoming ordinary across the rest of their lives. The public can be moralized into confusion, but not forever. Eventually convenience, usefulness, and pleasure begin eating the sermon from underneath.

Meanwhile, the institutions issuing the sermon will use the tools themselves. Of course they will. They would be deranged not to.

The record label that warns against AI will use AI to analyze hits, generate concepts, test markets, clean audio, explore hooks, and speed production. The studio that worries aloud about authenticity will use AI in development, editing, localization, effects, marketing, and audience analysis. The publisher that laments machine writing will use machine writing to draft catalog copy, summarize manuscripts, translate samples, and prepare campaigns. The platform that labels AI for the public will depend on AI internally for nearly everything it does at scale.

This is not hypocrisy in the dramatic sense. It is the normal metabolism of power. A tool is dangerous when it strengthens outsiders and prudent when it strengthens incumbents. The moral vocabulary changes according to who is holding the instrument.

The great comic element is that the campaign may be defeated not by argument, and not even by law, but by adoption. The more creators use AI, the harder it becomes to maintain a useful stigma around AI creators. At the beginning, a boundary can be drawn with a ruler. On one side are the approved makers, on the other the suspicious ones. But then the suspicious ones multiply. They include students, illustrators, editors, teachers, musicians, small businesses, podcasters, filmmakers, animators, marketers, game designers, and eventually the very companies that helped popularize the suspicion. The border does not get crossed in one dramatic march. It is crossed casually, daily, almost absentmindedly, until one morning everyone looks up and realizes the checkpoint is standing in an empty field.

This is how technological panics usually decay. Not with a confession. Not with a clean victory. Not with the old guard saying, “We were wrong, and the future may proceed.” They decay because the disputed tool becomes too common to isolate. Photography did not remain a scandal once everyone had a camera. Digital editing did not remain an outrage once every studio depended on it. The synthesizer did not destroy music; it entered music, changed it, and then became part of the furniture. The internet was once treated as an alien force corrupting publishing, journalism, commerce, childhood, politics, and attention. Some of that was even true. Yet the internet did not remain a separate moral category. It became the weather.

AI is headed toward the same condition. Not because every claim made for it is true. Not because every criticism is foolish. Much AI work is bad. Some uses are exploitative. Some disclosures are reasonable. Some legal questions are serious. But none of that rescues the larger distinction from its fate. A society cannot permanently divide creators into clean and unclean based on a tool that is becoming part of nearly every act of digital creation.

The lasting question will not be whether AI was involved. It will be what the human being was trying to do with it. That is a less convenient question for gatekeepers because it returns attention to judgment, intent, skill, taste, and result. It asks whether the work matters, not whether the workflow can be ritually purified. It treats tools as tools and creators as responsible for what they make with them.

That is where the argument is going, whether the current priesthood likes it or not. For a brief historical moment, it has been possible to make AI use feel like a confession. Soon it will feel like admitting that one used software. The accusation will not disappear because it was answered point by point. It will fade because it no longer points to a distinct class of people. The heretics will not storm the cathedral. They will become the congregation, the choir, the janitorial staff, and eventually the committee in charge of refurbishing the stained glass.

Then the label will lose its magic. The stain will not wash away. It will simply become too widely distributed to function as a stain. And the thing now treated as a threat to creation will take its place among the ordinary instruments of creation, where yesterday’s miracles always end up: not on the altar, not in the dock, but in the drawer with the scissors, the tape, the charger, and the spare batteries.

Copyright © 2026 by Paul Henry Smith

I don’t mean that as a moral criticism. Frankly, I understand it. If I were twenty-two years old, graduating into a world where every executive, dean, consultant, futurist, and LinkedIn foghorn had spent the past few years explaining that the thing I had just studied was about to be “transformed,” “disrupted,” “reimagined,” “augmented,” or otherwise run through the corporate smoothie machine, I might boo too.

Especially if the person saying it sounded delighted.

There is a certain kind of AI optimism that lands on young people the way a landlord lands on a tenant: technically smiling, already measuring the windows. It says: here is the future, it is inevitable, adjust yourself accordingly. It says: learn to use the tools. It says: deal with it. It says: your anxiety is noted, and now please clap for the rocket ship that may be aimed at your profession.

So yes, students booing AI is not irrational. It may be one of the more coherent responses available.

But I gave a talk at Oberlin Conservatory about AI and music, and no one booed.

I don’t say that because I heroically cracked the Gen Z anti-AI code, like some cardiganed hostage negotiator wandering into the quad with a MIDI keyboard. I say it because the experience clarified something I think is being missed in the public argument about AI, especially in the arts.

The question is not whether students love AI or hate AI.

The question is what kind of human activity AI is being invited into.

If you bring AI into a room as a way to replace effort, bypass skill, generate product, accelerate output, or flood the world with more disposable content, then yes, people who have devoted their lives to learning a craft may object. They may boo. They may boo with excellent diction.

But if you bring AI into a room as a way to deepen the human work, to make listening sharper, practice richer, comparison faster, curiosity more vivid, then the reaction can be completely different.

At Oberlin, I gave a talk called The Moment of Music: How Performers Create a Living Present, and What AI Can Help Us Hear. The premise was not that AI can now make music, therefore musicians should adapt. That is the dullest possible version of the conversation. The premise was that the deepest value of music was never merely the production of sonic artifacts in the first place.

Music happens in time. A sound appears, disappears, and is replaced by the next sound. If we experienced music as nothing more than one event after another, music would be impossible. But that is not what happens. Earlier sounds remain active. Future sounds become anticipated. The present expands. A phrase begins to mean more than the notes inside it. A cadence arrives not as a chord but as the fulfillment of something we have been carrying. A performance can make time feel larger than ordinary life usually allows.

That was the subject of the talk: not AI as output, but music as lived experience.

I used a few simple demonstrations. Three notes: G, B, D. Played one after another, they become a G major triad. But by the time the D arrives, the G and B are physically gone. They exist only because they remain active in perception. Then a children’s tune, Lightly Row, where the second little motive can be heard as lower only because the first little motive is still somehow present enough to compare it with. Then a repeated figure from Beethoven’s Sixth, artificially extended until the mind starts turning it into wallpaper. Then several AI-generated re-renderings of a phrase from Bach’s Double Violin Concerto: a traditional version, a beat-dominant version, a marching-band version, and a 1960s film-score version that was, honestly, kind of thrilling.

But the point was never, “Look what AI made.”

The point was, “Listen to what changes.”

What happens when the beat carries the present moment but does not align with the phrase’s deeper motion? What happens when a phrase’s energy, pulse, and harmonic direction all point the same way? What happens when a transformed surface actually makes the music’s direction easier to hear?

AI was not being used as a composer. It was not being used to generate yet another track for the great landfill of infinite content. It was being used as a comparison machine, a way of quickly placing musical possibilities next to one another so that human listeners could hear more sharply.

That is a completely different use case.

And it matched something else I saw at Oberlin that weekend. I attended a dance performance in which AI-generated music made with Suno had been used in the creative process and then brought into the world of live performance, with human musicians singing and playing. No one booed. The audience loved it.

Why?

Because the AI content was not the point.

The point was what the artists and performers did with it. The AI-generated material became part of a living process. It was a sketch, a provocation, a source, a scaffold, a strange visitor to the rehearsal room. Once human beings started shaping it, adapting it, singing it, playing it, embodying it, the AI-ness of the origin became almost beside the point. No one sat there thinking, “Ah yes, a generated artifact.” They were watching people make something happen.

That distinction matters enormously.

Most of the public argument about AI and music is still trapped inside the artifact model. Can AI make a song? Can it imitate a style? Can it generate a cue, a track, a chorus, a beat, an orchestral passage, a convincing vocal? Yes. Increasingly, yes. It can and will generate oceans of musical stuff. More tracks, more clips, more sonic wallpaper, more background music, more plausible imitations of things humans have already done.

But the world does not need more sound in the abstract. We are drowning in sound.

The deeper question is whether music can still create experiences that people actually need: experiences where attention gathers, time changes, people become more present, and a room feels briefly more coherent than ordinary life usually allows.

That kind of experience is not the same as an audio file. It is not the same as a prompt result. It is not the same as a recording, even a beautiful one. It happens when human beings are making sound in real time, listening, adjusting, carrying what just happened into what is about to happen, making the next sound feel not merely next but necessary.

This is what many young musicians actually care about. Not “content.” Not “deliverables.” Not sonic product. They care about the real thing: the reason music became too important to ignore in the first place.

And this is why the AI conversation so often goes wrong.

Students are not necessarily booing because they are technophobic. They are booing because they can feel, sometimes faster than the adults in the room, when a technology is being presented as a substitute for human development rather than a support for it. They can feel when “tool” is being used as a polite word for “replacement.” They can feel when the people selling inevitability have not bothered to understand what is being lost.

But that does not mean AI has no place in the arts.

It means the place has to be earned.

For me, the interesting uses of AI in music are not the ones that make more artifacts. They are the ones that help musicians become better musicians. Can AI help a performer hear five possible realizations of the same phrase? Can it help reveal how texture affects continuity, how decay affects phrasing, how pulse affects expectation? Can it help a student compare the difference between a line that merely continues and a line that actually moves? Can it generate examples strange enough to make us hear familiar music with fresh ears?

If so, that is not the atrophy of musicianship. That is a training tool.

Not a substitute for sweat. A way of directing the sweat.

The Oberlin students who wrote against chatbot use were right to defend “generative sweat.” That phrase is wonderful. Real learning requires friction. It requires effort, failure, repetition, memory, attention, and care. If AI becomes a way to skip that, then it is corrosive.

But sweat by itself is not sacred. People have sweated for stupid reasons across all of human history. The question is what kind of effort makes us more capable, more perceptive, more alive.

The best use of AI is not to eliminate effort. It is to make the right effort more visible.

That is what happened, in small ways, at Oberlin. AI was not treated as a vending machine for finished art. It was used to open a window onto musical questions. The human beings in the room still had to listen. They still had to judge. They still had to make meaning. They still had to decide what mattered.

And once the AI had done its useful little job, no one cared much about the AI-generated content itself.

That may be the most important part.

Once the musicians had interacted with it, learned from it, transformed it, or moved past it, the artifact did not retain some special halo. It was not the treasure. It was part of the process.

That is exactly where AI belongs if it is going to belong in music at all.

Not as the destination.

As a tool in the rehearsal room, the studio, the classroom, the composer’s sketchbook, the performer’s listening practice. A way of multiplying comparisons. A way of making hidden assumptions audible. A way of asking better questions faster.

The future will have plenty of generated music. It will have enough output to bury every platform on Earth under a glossy avalanche of plausible sound. Fine. Let it.

But output is not the same as musical life.

What matters is whether human beings can still gather around sound and make something happen that no artifact alone can produce: a present that expands, a room that listens differently, a moment where people feel directly that life can become more coherent, more shared, more awake than it usually feels.

That is the work.

And if AI can help musicians practice toward that, good. Let’s use it carefully, skeptically, creatively, and without surrendering the point.

But if AI is offered as a replacement for that work, then by all means, boo.

Boo with feeling.

Boo in tune.

Then go back to the practice room and make something real happen.

Well, let me start by saying this…

I get you.

I actually do.

All you anti-AI music people, you’re not crazy. You’re not villains. You’re not sitting there like some cartoon bad guy stroking a cat going, “Yes… let us crush creativity.”

No, you think you’re doing the exact opposite.

You’re sitting there going:

“Hey… this is messed up.”

You see these AI models, right?

You’re like:

“Hold on… they trained these things on our music??

Without asking?? Without paying??”

And you’re thinking about:

• the session musician who got $200 and a sandwich

• the indie artist grinding for ten years

• the producer who built a sound brick by brick

…and now some machine just absorbs all of it and starts spitting stuff out?

Yeah. That feels gross.

I get why your instinct is:

“No. Shut it down.”

“We need rules.”

“We need enforcement.”

“We need to stop this before it wipes everybody out.”

That instinct?

Totally human.

Totally understandable.

But here’s where it goes sideways.

Because what you think you’re doing is:

Protecting artists from exploitation.

What you are actually helping create is:

A system that controls who is allowed to create.

And those are not the same thing.

At all.

Let’s walk through what you want.

You want:

• AI detection

• Upload filtering

• Labeling

• Enforcement

• Payment if AI was used

Right?

Because in your mind, that leads to:

“If you used stolen data… you shouldn’t profit.”

Okay.

Stay with me.

Now let’s fast forward like… six months.

Not sci-fi. Not dystopia. Just… the next logical step.

• You upload a track.

• You made it yourself.

• You’re proud of it.

You used some tools—maybe a little AI-assisted EQ, maybe some generative texture thing, maybe you didn’t even realize it was AI because everything is AI now.

And the system goes:

“This contains AI-generated elements.”

You go:

“Okay… but it’s original.

It doesn’t copy anything.”

And the system goes:

“That’s not the question.”

That’s the shift.

That’s the part you didn’t sign up for.

Because in your head, the rule was:

“If it copies, it’s wrong.”

But the system you asked for? Doesn’t care about copying.

It cares about process.

Did you use the tool?

Yes or no.

And now suddenly, you’re not being judged on:

• what you made

• how original it is

• whether it infringes anything

You’re being judged on:

• how you made it.

And that is a completely different world.

Because once you move the line there… once you say:

“Using this tool creates an obligation…”

You’ve just given whoever controls that tool—or claims ownership over its training—the ability to say:

“Anything made with it? We get a piece.”

Even if your work is completely new.

Even if it violates nothing.

Even if it’s better than anything they’ve ever made.

And here’s the part that should hit you in the gut.

The exact system you’re asking for to stop exploitation…

is the perfect system to enforce it at scale.

Because now:

• Platforms have to comply

• Creators have to prove innocence

• Labels don’t have to prove infringement

They just go:

“Hey… that tool?

Yeah, that traces back to our catalog.

So we’re involved now.”

That’s it.

No courtroom.

No melody comparison.

No “this bar matches that bar.”

Just:

“You used it. Pay us.”

And if you don’t?

What are you gonna do?

Fight them?

With what money?

With what legal team?

You’re gonna do what everybody does.

You’re gonna go:

“Alright… what’s the fee?”

And now we’ve arrived.

You started here:

• “We need to protect artists from being exploited.”

And you ended here:

• “Artists must pay to create.”

That’s the inversion.

That’s the trap.

And the reason it works (the reason it’s so sneaky)

is because it feels righteous the whole way through.

At no point do you feel like you’re doing something wrong.

You feel like you’re defending fairness.

You feel like you’re standing up for human creativity.

Meanwhile, the people who actually benefit?

They don’t argue.

They don’t correct you.

They don’t go:

“Hey… just so you know, this logic is gonna boomerang.”

They just go:

“Yeah. Keep going. You’re doing great.”

Because they know something you don’t.

They know that once the rule becomes:

“Tool used = payment owed”

It doesn’t matter who the artist is anymore.

It only matters:

Who owns the tool.

And spoiler alert:

That’s not you.

So yeah.

Be angry about training data.

Ask hard questions.

Demand fairness.

But be very, very careful about what you ask for in response.

Because if you get exactly what you want…

you may find that the system you built to protect yourself…

is the one that quietly decides…

You don’t get to create for free anymore.

Copyright © 2026 by Paul Henry Smith

I stood at the edge of the bay and looked across the water—not with any fixed intention, but with that idle steadiness which sometimes overtakes a man when the world before him appears sufficiently distant to require no immediate response. The far shore lay stretched in a dim, patient line, and rising from it—thin as a needle drawn lightly across the sky—was the carillon tower of the university. It was so small at that distance that one might almost overlook it entirely, mistaking it for some trivial irregularity in the horizon. The greater presence was not the land, nor the structure, but the sky itself, which spread in a vast and untroubled expanse, crowding out all lesser claims.

I considered, after a time, that the tower must be near ten miles away. It was not an extraordinary distance; a man might traverse it in an hour by car, or in several on foot, yet already it rendered what was solid and defined into something nearly unreal. If a man stood within that tower—if he leaned against its railings, if he lifted his hand, if he spoke—none of it would reach me. He would exist, certainly, but not for me. The air would take him, the light would dissolve him, and I would be left only with that faint vertical suggestion, which might at any moment vanish into the sky that dwarfed it.

I had, not long before, been reading of a man whose name appeared with such frequency that it seemed, for a time, to occupy more space than others. His movements were noted, his remarks repeated, his presence treated as though it exerted a certain pressure upon the world.

Yet if he stood now in that tower across the bay—if he raised his hand or turned his head—I would not see him. Whatever magnitude he possessed did not extend this far. At such a distance, he would be indistinguishable from anyone else, or from no one at all.

The thought did not come to me as a criticism. It arrived instead with a kind of simple finality, as though I had observed something obvious that had previously been obscured by habit. The world, when seen at even a modest remove, did not preserve these distinctions. It reduced them. It leveled them without effort.

And distance, I began to suspect, was only the lesser force at work.

Time was the greater.

There was a period, long before any tower had been raised, before any name had been spoken or recorded, when the land was inhabited by little more than moss. It spread itself across the bare surfaces of the earth with no apparent ambition, no declaration of purpose. It did not seek to be remembered; it did not attempt to impress itself upon anything beyond its immediate reach. It simply persisted.

And it did so for forty million years.

The number, when first encountered, presents itself as a curiosity. When considered, it grows unwieldy. It extends beyond the ordinary grasp of imagination, until it becomes less a quantity than a condition—a vast, unbroken duration in which the quiet, unassuming presence of that life endured without interruption.

We, by contrast, have been here in our present form for scarcely two million years. Eight billion of us now move upon the earth, filling it with motion, with speech, with endless assertions of our own significance. We build, we measure, we proclaim. We set our names upon things and call them ours.

And yet, when placed beside that earlier age, we are brief.

It seemed to me, standing there, that we would not last as long as moss.

This conclusion did not arise from despair, nor from any particular dissatisfaction with the present. It was simply an acknowledgment of proportion. The world had endured states vastly different from this one, and it would endure others still. Our presence, however vivid it might seem from within, did not guarantee its own continuation.

The wind moved lightly across the water. A gull passed overhead, its motion steady and unremarked. The tower remained where it was, though I knew that even it, in time, would not.

I found no urgency in these thoughts. They did not press upon me with any demand for resolution. They settled instead into a quiet understanding, one that neither troubled nor consoled, but merely clarified.

It was with this clarity still upon me that I turned from the bay and made my way into the city, where the streets received me with their usual indifference. The crowds moved as they always had—swiftly, purposefully, each individual enclosed within the narrow circle of their own concerns. I entered among them and was carried along without resistance.

There were the usual small adjustments of motion: a step taken to avoid another, a slight turning of the shoulders, a glance that flickered and was gone. These were not recognitions. They were corrections, made unconsciously, without record.

I passed through.

In a coffee shop, I joined a line that advanced in small increments toward the counter. The room was filled with the low machinery of daily exchange: the hiss of steam, the dull contact of ceramic against wood, the subdued murmur of voices that neither rose nor fell enough to be distinguished. Names were called and vanished into the air almost as soon as they were spoken.

When my turn came, I stepped forward.

“Black coffee,” I said.

The man behind the counter nodded without looking at me. His hands moved. The cup was filled, the transaction completed, and I was already being replaced by the next figure in line. There was no rudeness in this. It was simply the manner of things.

I took the cup and moved aside. The lid was not properly secured; I adjusted it. Such minor corrections, I had observed, often fell to the one who noticed them.

For a moment, I considered whether to say something more—to add a word, a remark, anything that might extend my presence in that exchange beyond its immediate necessity. The impulse was faint and without conviction. I let it pass.

I seated myself by the window.

Outside, the movement continued uninterrupted. Inside, the same pattern repeated itself with quiet efficiency. Nothing lingered. Nothing adhered.

After some time, another person entered and took a seat near mine. There was nothing at first to distinguish them, yet there was a slight hesitation in their manner, as though their actions did not proceed with the same easy continuity as those around them. They opened a laptop, then closed it. Their hands rested idle. A breath escaped them, not loudly, but with a certain weight that suggested it had not been intended for the room.

Without turning, they spoke.

“May I ask you something?”

The question was delivered cautiously, as though it might not be welcome.

“You may,” I said.

They paused briefly, as if arranging their thought not for effect, but for accuracy.

“Do you ever feel,” they said, “as though you move through everything without really touching it?”

The words were plain, yet they carried with them a precision that left little room for misunderstanding.

“Yes,” I said.

They turned then and looked at me directly. There was no surprise in their expression, only a quiet recognition, as if something long suspected had at last been confirmed.

“I thought it was only me,” they said.

“No.”

We sat in silence for a time. It was not the empty silence of strangers, but something more deliberate, as though the fact that had been spoken required no immediate elaboration.

After a while, they asked, “What do you make of it?”

“Very little,” I said.

They considered this.

“You are not troubled by it?”

“Not now.”

“How did that come about?”

I looked out at the street, where the steady procession of figures continued without pause.

“I ceased expecting it to be otherwise,” I said.

They were silent for a moment.

“And that is enough?”

“It is sufficient.”

They lowered their gaze, then lifted it again.

“Then what remains?” they asked.

I turned back to them.

“Finding those who understand it,” I said.

They studied me with a thoughtful expression.

“And this—” they said quietly, indicating the space between us—“is that what this is?”

“Yes.”

At that, something in their face eased—not greatly, but enough to be seen. It was the expression of one who has discovered that a private condition is not, in fact, solitary.

We spoke little after that. There seemed no need.

When at last I rose to leave, they did the same. We parted without ceremony, each returning to the larger movement from which we had briefly diverged.

Outside, the street received me again. The crowd closed around me. The world proceeded as it had before, unchanged by anything that had passed.

And yet, as I walked, the thought came to me with a quiet completeness, as though it had been forming long before I had the words for it.

I understood then what I was.

Not in any grand or final sense, but in the simple, descriptive way one might understand the nature of a thing that has long been observed.

I was, in the world of men, something like those particles I had read of—moving always, present everywhere, and yet leaving almost no trace of passage. I crossed paths, I entered and exited, I was there and then not there, and in most cases nothing bore the mark of my having been.

A neutrino.

The word settled into place without resistance.

It did not diminish me. It did not elevate me. It merely named the condition.

And within that condition, I had found, however rarely, another of the same kind—one whose path, for a moment, did not simply intersect mine and vanish, but was recognized.

It was a small thing. It would not endure. It would not alter the course of the world, nor extend the brief tenure of our kind upon the earth.

But it was not nothing.

Copyright © 2026 by Paul Henry Smith

Jordan

[Orchestral overture]

Imagine a new technology drops today, right?

And the government immediately moves to ban it.

They claim it’s going to fundamentally corrupt the youth and cause the absolute collapse of the state.

You’d probably think it was, I don’t know, a biological weapon.

Or maybe some kind of unregulated neuroimplant.

Alex

Exactly.

But if you rewind to about 380 BCE, Plato was making that exact argument about a new type of flute.

It is just a stunning historical reality.

We tend to think of the history of music as this upward trajectory of universal celebration.

Jordan

Right, where society just marvels at the next great masterpiece or a cool new instrument.

Alex

Yeah, but if you look at the primary sources, the reaction to new musical expression is almost always sheer, unadulterated terror.

Jordan

Which is exactly what we are getting into today.

Welcome to The Deep Dive.

Our mission today is to track the overarching through-lines of this fear.

We want to figure out why new music and new music tech always seem to terrify society.

And what’s uniquely different about the panics you see in your social feeds today versus what’s exactly the same.

And what conclusions we can draw about the future of human expression.

Okay, let’s unpack this.

Alex

The most striking realization from this research is that while the target of the panic constantly evolves, shifting from ancient lyres to 19th-century ballroom dances to 2026 AI track generators, the underlying rhetoric remains shockingly consistent.

It’s basically the same script every time.

Jordan

It really is.

To understand the AI anxiety we’re living through right now, we have to look at how early societies viewed music.

They didn’t see it merely as an art form.

They saw it as a highly dangerous technology of the physical body.

Alex

Let’s explore that, because the level of state control over a melody in antiquity is wild.

You mentioned Plato warning that musical innovation leads to lawlessness.

Jordan

Oh yeah. He thought it was a direct threat to the state.

Alex

But it wasn’t just a Western phenomenon.

In early Confucian statecraft, there was a massive push to banish the regional music of Zheng.

Jordan

Right, because it was classified as lewd.

Alex

Exactly. It was treated like a political hygiene issue.

Imagine the government banning a Spotify playlist because they genuinely believe it’s a threat to national security.

Jordan

It sounds absurd now, but as history progresses, that fear transitions into a fear of music corrupting the soul.

Which brings us to the religious panic.

Alex

If you read Augustine of Hippo, he agonizes over his own physical reactions to music.

Jordan

He felt guilty just for reacting to a song?

Alex

Totally. He felt like a criminal because he was more moved by the singing than the religious message.

Jordan

That’s incredible.

Alex

And it escalates.

Figures like John Chrysostom and later Puritan clergy framed dancing as a direct portal to evil.

Jordan

The Puritans did not mess around with dancing.

Alex

Not at all. Increase Mather literally described it as a devil’s procession.

Jordan

And then by 1816, the waltz is causing panic in London.

Alex

Yes, it was called an indecent foreign contagion.

Jordan

Because people were touching.

Alex

Exactly. That same anxious gaze appears again with the hula in the 1820s.

Missionaries framed it as morally disruptive and socially dangerous.

Jordan

It really does feel like they treated music as a kind of malware.

Alex

That’s exactly the pattern.

The state or church is the operating system, and new music is treated like a virus that hacks the body.

Jordan

That brings us to something the sources call “demonology by metaphor.”

Alex

Right. It’s about externalizing agency.

Instead of saying “I like this,” people say “the music is making me do it.”

Jordan

So the music becomes the villain.

Alex

Exactly. It absolves the listener of responsibility.

Jordan

But in the 20th century, the language changes.

Alex

Yes. The panic becomes scientific.

Ragtime was described as a public health issue.

Jazz was said to “demoralize the brain.”

Jordan

And those claims were often wrapped in racialized pseudoscience.

Alex

Exactly.

And that continues into rock and roll, where the focus shifts to physical behavior and neurological harm.

Jordan

Which leads us to the PMRC era.

Alex

Yes. The rhetoric becomes statistical moralism.

Explicit lyrics were linked to social epidemics like violence and suicide.

Jordan

So taste becomes framed as measurable harm.

Alex

Exactly. It transforms opinion into urgency.

Jordan

Then we get the machine panic.

Alex

John Philip Sousa warned in 1906 that mechanical music would destroy the human soul.

Jordan

Which sounds exactly like modern AI critiques.

Alex

It’s the same argument.

Later, unions protested synthesizers, fearing job loss.

Jordan

Which gets reframed as protecting culture.

Alex

Exactly. Economic anxiety becomes moral concern.

Jordan

Then we enter the digital era.

Alex

Yes. The panic moves into the legal system.

Home taping was “killing music.”

Sampling cases invoked biblical language.

Jordan

“Thou shalt not steal” in a court ruling is wild.

Alex

And then Napster and file sharing escalate everything.

Jordan

The industry calls users pirates.

Alex

Yes, turning consumers into criminals.

Jordan

But none of it stops the technology.

Alex

No. It just delays adaptation.

Jordan

Which brings us to today.

Alex

The authenticity crisis.

AI is framed not as corrupting us, but as replacing us.

Jordan

That’s the shift.

Alex

The fear is now an ontological insult.

Jordan

Meaning?

Alex

The fear that human creativity isn’t unique.

That it can be replicated.

Jordan

That’s a very different kind of panic.

Alex

Yes, but the pattern remains the same.

Panic, litigation, normalization.

Jordan

And eventually, integration.

Alex

Exactly.

Jordan

So what’s the takeaway?

Alex

Moral panics over music are rarely about the music itself.

They’re about power.

Economics.

Control.

And who gets to define authenticity.

Jordan

Every terrifying new technology eventually becomes just another tool.

Alex

Which leads to two questions you should always ask.

Jordan

Who is losing money?

Alex

And who is losing control?

Jordan

And maybe one more.

If machines can imitate everything…

Alex

What happens when there’s nothing left to imitate?

Jordan

Maybe the future of rebellion is just humans being gloriously imperfect.

Alex

Messy, offbeat, unmistakably human.

Jordan

Let’s hope so.

Thanks for joining us on The Deep Dive.

Until next time.

I did not see this immediately.

Like many beginners in this new world, I assumed that when the AI and I were working on something together, the correct response to difficulty was perseverance. Something breaks, you explain the problem again. The AI tries another fix. Something else breaks. You explain it a little more clearly this time, perhaps with the patience of someone speaking to a very intelligent but slightly distracted dog.

And sometimes this works. But sometimes it does not.

Sometimes you find yourself several hours into a project that now consists of files talking to other files, functions calling functions whose purpose neither of you entirely remembers, and a growing sense that every attempt to correct the system merely encourages it to invent a fresh and more creative form of malfunction. The AI apologizes with touching sincerity. It proposes improvements with the enthusiasm of a child who has just repaired the toaster by filling it with glue.

You try again.

After all, this is what software people do. They fix things. They untangle the wires. They trace the dependencies. They rebuild the architecture.

The difficulty, in my case, is that I am not one of those people.

Presented with a codebase of several thousand lines — none of which I wrote and all of which now seem to refer to one another in ways that feel both intimate and accusatory — I do not possess the tools required to diagnose its ailments. The idea that I will calmly step through the modules and identify the flawed assumption lurking deep within the structure is charming, but unrealistic, in the same way it would be unrealistic for me to repair the carburetor of a lawn mower by staring at it thoughtfully.

The usual engineering instinct at this moment is to persist.

This is actually not irrational. For most of the history of software, abandoning a codebase was an act bordering on madness. Once something existed, you had already invested an enormous amount of time and money into it, and starting again meant repeating that entire process from the beginning. The cost of starting over was too high.

This is how the world ended up with banking systems written in COBOL, a language whose cultural heyday coincided with the introduction of the color television.

But AI development changes this calculation in a peculiar way.

Because starting again does not necessarily mean months of labor. Sometimes it means only an hour.

When I reach that point — the moment when the AI and I appear to be circling the same misunderstanding like two polite strangers attempting to pass one another in a narrow hallway — I now do something that would probably cause a traditional engineering team to make that tense sucking through clenched teeth sound and wince.

I stop.

I gather the artifacts of the previous attempt.

The codebase.

The entire chat history.

The notes I took along the way.

A plain-language explanation of what I was trying to build, for whom, and why.

If I have any sense of what went wrong — “it kept crashing when X,” or “the timing logic seems confused” — I include that too.

Then I give all of this material to ChatGPT, asking it to assemble the sort of documentation an actual engineering and design team might produce: the context, the description of the problem, the requirements, the PRD, the tasks. It gives me a tidy packet of understanding, as if a group of sober professionals had spent several meetings deciding what the thing ought to be.

I take that bundle of explanation and hand it to a different coding AI. Or the same one, but a new project.

Fresh project. Fresh context.

No memory of the previous attempt’s peculiar habits.

And what happens next is occasionally remarkable.

The new system reads the materials, considers the problem, and produces something that looks suspiciously like a functioning implementation.

Not because the first attempt was useless, but because it served as a rehearsal. The earlier effort clarified the idea, exposed the weaknesses, and revealed assumptions that had not been visible before. And yet, despite some fatal flaw lurking in there, the second AI coding assistant could handle it.

I discovered this most dramatically while building a music production tool that began life, like so many AI projects, as a web application. That was simply what the system produced when asked to make a tool: a react app.

But after several attempts to dig out of some signal processing hole, I realized something quietly obvious. The people who would actually use this tool were not sitting around opening web apps. They were working inside music production software. Which meant the correct form for the tool was not a web application at all, but an audio plug-in — something that lives directly inside the digital workstation where they spend their time.

Did I attempt to convert the web application into that?

I did not.

I began again.

New instructions.

New architecture.

New AI.

The entire sequence — the failed attempts, the reconsideration of the idea, and the final working version — took roughly a day.

Which is when I began to recognize the situation for what it was.

Occasionally, without realizing it, you find that the elaborate thing you have been patiently constructing is not turning out like you imagined it would.

When that happens, you have been making an Amblongus Pie.

And at that point, the next step becomes very clear. Finish the recipe!

To Make an Amblongus Pie

(from “Nonsense Cookery,” in Nonsense Songs, Stories, Botany, and Alphabets, 1871, by Edward Lear)

1. Take 4 pounds (say 4½ pounds) of fresh Amblongusses, and put them in a small pipkin.

2. Cover them with water, and boil them for 8 hours incessantly; after which add 2 pints of new milk, and proceed to boil for 4 hours more.

3. When you have ascertained that the Amblongusses are quite soft, take them out, and place them in a wide pan, taking care to shake them well previously.

4. Grate some nutmeg over the surface, and cover them carefully with powdered gingerbread, curry-powder, and a sufficient quantity of Cayenne pepper.

5. Remove the pan into the next room, and place it on the floor. Bring it back again, and let it simmer for three-quarters of an hour. Shake the pan violently till all the Amblongusses have become of a pale purple colour.

6. Then, having prepared the paste, insert the whole carefully; adding at the same time a small pigeon, 2 slices of beef, 4 cauliflowers, and any number of oysters.

7. Watch patiently till the crust begins to rise, and add a pinch of salt from time to time.

8. Serve up in a clean dish, and throw the whole out of the window as fast as possible.

Copyright © 2026 by Paul Henry Smith

Over the past few years the AI ecosystem has been assembling itself into layers.

First came the models. Then came the tools that allow those models to interact with the world. Now we’re beginning to see protocols for agents to communicate with each other and frameworks that help orchestrate their work.

This stack is forming quickly. And like most technology stacks, it’s forming somewhat organically—different groups solving different problems, one layer at a time.

But when you look closely at the emerging architecture, a small question begins to nag.

Something may be missing.

Not another model.

Not another framework.

A way to describe the work itself.

Something interesting seems to be happening in the AI ecosystem.

Not just with models, which continue to improve at a remarkable pace, but with the infrastructure forming around them. A growing number of tools, frameworks, and protocols are appearing that help AI systems interact with the outside world, communicate with one another, and coordinate work.

When technologies mature, their surrounding infrastructure tends to organize itself into layers. It’s not usually planned this way. It simply emerges as different problems become visible and different groups start solving them.

Right now, the emerging AI stack appears to be settling into a few recognizable layers. They are still forming, and the boundaries are not yet sharp, but a rough picture is beginning to appear.

Looking at that picture, one question starts to nag.

There may be a layer missing.

Tool Access

The first thing AI systems need is access to the world.

Large language models are powerful reasoning engines, but by themselves they are sealed off from external information and capabilities. To do anything useful in practice they need ways to read files, query databases, call APIs, run code, or interact with internal systems.

A number of efforts are beginning to standardize this layer. One example is the Model Context Protocol (MCP), which provides a way for tools and data sources to expose capabilities that models can call.

In its simplest form, the interaction looks something like this:

AI → tools
AI → data
AI → services

The goal of this layer is straightforward: make it easy for AI systems to reach out and interact with real systems.

In traditional computing terms, this is roughly analogous to the I/O layer. It connects the reasoning engine to the environment in which work actually happens.

Agent Communication

Once multiple AI systems exist, another need quickly appears.

They need to talk to each other.

In many real scenarios, different AI systems will have different capabilities, access to different tools, or responsibility for different parts of a workflow. Allowing them to communicate makes it possible to distribute work across them.

This has led to emerging work around AI agent communication protocols, including efforts in the W3C community group exploring agent interoperability.

At a high level, this layer defines things like:

• how agents identify themselves

• how they describe their capabilities

• how they send tasks or requests

• how they return results

In simplified form, the interaction looks something like this:

Agent A → message → Agent B
Agent B → response → Agent A

If tool protocols give agents the ability to interact with the world, agent communication protocols allow them to interact with each other.

Orchestration

Once agents can call tools and talk to each other, another problem emerges.

Someone—or something—needs to decide who does what.

Today this role is mostly filled by orchestration frameworks. Examples include systems like LangGraph, AutoGen, CrewAI, or the planning components in platforms like Semantic Kernel.

These frameworks help coordinate workflows across agents and tools. They typically provide mechanisms for things like:

• task decomposition

• agent coordination

• tool invocation

• result aggregation

But each framework tends to define its own internal structures for describing work. A task in one framework may not look anything like a task in another.

And that observation leads to an interesting question.

What Is the Unit of AI Work?

When humans collaborate, work tends to be organized around tasks with purpose.

Someone might investigate a production anomaly, write a report, perform an analysis, or run an experiment. Each of these activities has a goal, a set of constraints, and an expected outcome.

AI systems increasingly participate in the same kinds of activities. But there is currently no shared abstraction for representing them.

The ecosystem has:

• tools

• agent communication

• orchestration frameworks

But there is no common way to describe the work itself.

Not a prompt.

Not a tool call.

Something higher level.

A convenient word for that might be a mission.

Missions as Structured Work

A mission could be thought of as a structured description of purposeful work.

For example:

Mission
  goal: Diagnose checkout conversion anomaly
  constraints:
    - read-only analytics
    - no customer data export
  budget:
    - compute: $5
    - time: 10 minutes
  output:
    - diagnostic report

This object describes several things at once.

It captures the intent of the work: diagnosing an anomaly.

It specifies constraints: what data may be accessed and what actions are forbidden.

It defines resources: time and compute limits.

And it describes the expected outcome.

Agents and tools can then operate within this context as they attempt to achieve the mission.

Importantly, missions are not something agents invoke in a strict hierarchy. Rather, they define the environment within which agents operate.

Separating Reasoning from Authority

One of the motivations for this kind of structure is the separation of two responsibilities that are often mixed together in early agent systems.

Agents are very good at reasoning about problems. They can propose plans, generate ideas, and explore possible solutions.

But deciding what actions are allowed is a different problem.

A mission could include a policy envelope that defines boundaries for execution. For example:

policy:
  allowed_data:
    - analytics
    - logs
  forbidden_data:
    - customer_pii
  execution_limits:
    max_steps: 100
    max_parallel_agents: 5
  side_effects:
    allowed:
      - sandbox execution
    forbidden:
      - production changes

Under this architecture:

• agents propose actions

• policies determine whether those actions are permitted

This pattern already appears in many large systems. Operating systems separate processes from permission systems. Cloud infrastructure separates services from IAM policies. Kubernetes separates workloads from RBAC rules.

The same separation may eventually prove useful for AI-mediated work.

Missions as Living Objects

Another helpful feature of missions is that they can have lifecycle state.

Instead of simply starting a workflow and hoping it finishes correctly, the system can track the mission as it progresses.

For example:

submitted
admitted
active
paused
completed
failed
cancelled

This allows the system to answer operational questions such as:

• Is the mission still running?

• Has it exceeded its time or compute budget?

• Was it cancelled by a human?

In practice there might also be a distinction between mission definitions and mission runs, just as workflow systems distinguish between templates and individual executions.

Avoiding the “Jungle Soldier” Problem

There is an old story about soldiers stationed on remote islands during World War II who continued carrying out their wartime duties years after the war had ended.

The problem was simple: they had never received the signal that the mission was over.

AI workflows can encounter a similar failure mode if they are implemented as rigid sequences of steps. A plan created at the beginning of a process may no longer make sense as conditions change.

A more resilient design borrows a pattern that appears in modern infrastructure systems such as Kubernetes: the reconciliation loop.

Instead of blindly executing a fixed sequence of steps, the system repeatedly asks:

• Is the mission still relevant?

• Has the goal already been achieved?

• Has the environment changed?

This makes the system adaptive rather than brittle.

Congestion

As the number of agents grows, another issue begins to appear.

Even well-behaved agents can collectively overwhelm shared resources.

Agents may spawn subtasks, launch analyses, call APIs, or generate additional work. When many agents do this simultaneously, the system becomes congestible.

The early internet experienced something similar before congestion control mechanisms were introduced. The problem was not malicious senders. It was simply many independent actors sharing the same network.

TCP eventually solved this problem for packet transmission.

Large AI ecosystems may encounter an analogous challenge for work itself.

Mission Admission

One way to address this is to separate mission proposal from mission admission.

Agents might be allowed to propose missions automatically. But a control layer would determine which missions are actually admitted into the system.

Conceptually:

Agent → propose mission
↓
Mission control evaluates
↓
admit / delay / deny

This layer could enforce things like:

• rate limits

• budgets

• approval policies

• system load constraints

Without such mechanisms, automated agents could easily create runaway workloads.

Parent Missions

If missions can create other missions, another structure becomes useful: parent-child relationships.

For example:

Parent mission
├─ child mission
├─ child mission
└─ child mission

Child missions inherit constraints from their parent. They may share budgets or policy envelopes.

This prevents the system from becoming unbounded. Otherwise a parent mission could evade limits simply by spawning unlimited children.

A Mission-Centric Architecture

Taken together, these ideas produce a system that might look roughly like this:

Missions define the work.

Agents determine how to pursue it.

Tools provide the capabilities required to execute it.

Why This Pattern Might Appear

Coordination layers tend to emerge when systems reach a certain scale.

• Railroads developed dispatch systems.

• Electrical grids developed load balancing.

• The internet developed congestion control.

Each solved the same underlying problem: many independent actors sharing a common environment.

AI ecosystems are beginning to exhibit similar dynamics.

Two Possible Futures

At this point it is not clear how the ecosystem will evolve.

One possibility is that each vendor platform develops its own mission systems internally. Coordination remains platform-specific, and interoperability happens only at lower layers.

Another possibility is that shared primitives gradually emerge for describing and governing AI work. Different orchestration engines could still exist, but they would operate on a common understanding of missions.

History suggests that coordination layers often appear eventually.

The question is rarely whether they appear.

The question is whether they emerge deliberately, or only after fragmentation makes them necessary.

A Small Linguistic Aside

The AI field has introduced a number of creative phrases in recent years.

“Agentic AI” has become a widely used term.

If mission architectures ever become common, someone will inevitably suggest another phrase:

Missionic AI.

Whether the English language is ready for that particular invention remains an open question.

For now it may be enough simply to talk about AI systems operating under missions.

Some Thoughts

As AI systems multiply and begin generating work for one another, the nature of the challenge shifts.

The question stops being:

What can these systems do?

It becomes:

How do many such systems operate together without overwhelming the environment they share?

Whether the answer ultimately takes the form of mission layers, schedulers, policy systems, or something else entirely remains to be seen.

But the conditions that historically produce coordination layers are beginning to appear.

And that makes the question worth exploring.

Open Questions

Discussions about new architectural layers tend to provoke a particular kind of reaction from engineers. The instinct is usually not excitement but interrogation.

That instinct is healthy. Systems become reliable precisely because people keep asking uncomfortable questions.

Several questions tend to arise quickly when the idea of a “mission layer” is introduced. Some of them are worth addressing directly.

Couldn’t Agents Just Create Missions All the Way Down?

One obvious concern is that if missions become first-class objects, then agents could simply start creating new missions endlessly.

An agent assigned a task might decide it needs sub-missions. Those sub-missions might spawn further missions. Very quickly the system might resemble a kind of recursive explosion of work.

This concern is closely related to something well known in operating systems: the fork bomb. A program repeatedly spawns new processes until the system collapses under its own activity.

If missions can be generated programmatically, what prevents the same outcome?

The key distinction is between proposing a mission and admitting a mission into the system.

An AI agent could certainly propose new missions. But that does not necessarily mean those missions automatically execute. In many large systems, work requests pass through an admission stage before they become active.

For example:

• Operating systems allow programs to request new processes, but schedulers control how they run.

• Databases accept queries, but planners and quotas determine how resources are allocated.

• Kubernetes allows workloads to be defined, but admission controllers and schedulers determine whether they run.

A mission architecture might follow a similar pattern.

Agents could propose missions as part of their reasoning process, but a mission admission layer would decide whether those missions are accepted, delayed, or rejected.

This separation between proposal and admission helps keep recursive behavior bounded.

Could AI Just Automate Mission Creation Anyway?

Another variation of the same question is slightly subtler.

Even if humans are the ones who originally define missions, AI systems could easily automate the creation of many more missions downstream. If agents are allowed to create missions, what prevents them from flooding the system?

In practice, most large systems address this through budgets and quotas.

A mission might carry limits such as:

• maximum number of subtasks

• maximum compute budget

• maximum number of child missions

Child missions could inherit these constraints from their parent. That way an entire mission tree shares the same overall resource envelope.

For example, a research mission might have a budget of ten compute-hours. Whether the work is performed by one agent or ten subordinate missions, the total consumption cannot exceed that budget.

This approach prevents recursive expansion from becoming unbounded.

Doesn’t TCP Already Handle Congestion?

Another common reaction is that modern networks already solve congestion problems. TCP, for example, regulates packet flow across the internet. Why would AI systems need additional coordination layers?

The answer is that TCP regulates packet transmission, not task creation.

TCP ensures that senders do not overwhelm routers with packets. It does not prevent thousands of independent programs from deciding to send work to the same API, database, or compute cluster at once.

In other words, TCP controls the speed of traffic on the road. It does not decide how many vehicles are allowed to drive toward the same destination.

Large AI systems may encounter congestion at a different level: not packets, but work itself. That kind of congestion may require different coordination mechanisms.

Isn’t This Just Reinventing Workflow Systems?

Some readers may recognize similarities between missions and traditional workflow engines. Tools like Airflow, Temporal, and BPMN systems already coordinate complex tasks across many components.

The difference lies in how the work is described.

Traditional workflows typically specify a fixed sequence of steps. They describe how work should be executed.

A mission describes intent and constraints, but not necessarily the exact sequence of actions required to achieve the goal. Agents may plan different approaches depending on what they discover along the way.

In that sense, a mission resembles a goal specification, while the plan to achieve it may evolve dynamically.

Who Controls the Mission System?

Another natural question concerns governance.

If a mission system coordinates large amounts of AI activity, where does it live? Who controls it?

There is no single answer.

In practice, several possibilities might coexist:

• enterprise environments might operate their own mission control systems internally

• vendor platforms might implement mission governance inside their infrastructure

• open standards might eventually define interoperable mission descriptions

The internet itself evolved in a similar way. Individual networks operated independently while agreeing on shared protocols.

A mission layer, if it ever emerges, might follow a similar path.

Do We Even Need This?

Perhaps the most important question is the simplest one.

Is any of this actually necessary?

The honest answer is that no one knows yet.

Many coordination layers in computing were not designed in advance. They emerged gradually when systems reached a scale where uncoordinated activity became unstable.

Right now AI ecosystems are still relatively small. But if the number of agents grows dramatically, and if those agents begin generating work for each other, some form of coordination may eventually become necessary.

Whether that coordination takes the form of missions, schedulers, policy systems, or something entirely different remains an open question.

For now, the goal is not to prescribe an architecture. It is simply to notice the pattern beginning to appear, and to ask whether it might eventually need a name.

Copyright © 2026 by Paul Henry Smith

Oh really? Now we’re worried about homogenization?

Where the hell have you people been for the last forty years?

I love these tech panic articles. They always act like humanity was this wild jungle of creativity until ChatGPT showed up with a clipboard and started issuing personality citations.

No. We’ve been sanding the edges off everything since the invention of the focus group.

Take music. Remember the loudness wars? That glorious era when every album had to be louder than the last one until the waveform looked like a solid brick? Engineers just smashing compressors like they were tenderizing a steak.

Then rhythm flattened out too. Every pop song now: boom clap boom clap boom clap. Four-four time forever. Beethoven’s somewhere in the afterlife going, “Did nobody bring the odd meters? I left a whole bag of them!”

Car design? Same story. You look at the road today and it’s like somebody spilled a tray of graphite-colored soap bars onto the freeway. Black, white, silver, gray. Every car looks like a bar of tech startup hand soap.

When I was a kid cars looked insane. Lime green. Burnt orange. Purple so aggressive it looked like Prince personally approved the paint job. You’d see a Cadillac the color of radioactive taffy and think, “Yeah, that guy definitely sells fireworks out of a van.”

Now everything looks like it was emotionally sedated by a design committee.

And food! Every city on earth now has the exact same restaurants.

Burger place.

Ramen place.

Artisanal pizza place.

Third-wave coffee place where the barista explains the childhood of the bean.

You could parachute blindfolded into Lisbon, Chicago, or Singapore and within ten minutes you’d be eating a $19 sandwich with truffle oil and a personality crisis.

But the one that kills me is the YouTube accent.

You know the voice.

“Hey guys what’s up guys today we’re gonna talk about…”

Where are you from? Nebraska? Sweden? Mars? Nobody knows! It’s the International Influencer Dialect. A cheerful, neutral tone designed to offend absolutely no one on Earth.

And now suddenly researchers are like, “AI might standardize how people speak and think.”

Yeah. That’s because the thing is literally a statistical averaging machine. It predicts the most probable word sequence based on mountains of existing text. Of course it sounds like the middle of the bell curve. That’s what you trained it to do.

But humans have been chasing the middle of the bell curve since marketing departments discovered the phrase “broad appeal.”

Look at fashion. Everybody talks about individuality and then you walk outside and the entire species is dressed like they’re about to hike a polite mountain.

Gray hoodie.

Neutral sneakers.

A jacket that looks like it was designed by a weather app.

But nothing illustrates the species better than a hotel breakfast buffet.

You ever go to one of those things?

That is the United Nations of Bland.

Every culture on Earth represented… and somehow the food tastes like it was emotionally processed by a therapist.

There’s always those scrambled eggs that have the texture of wet packing peanuts.

The sausage that tastes like it studied meat in a book once.

And then the waffle machine.

You know the one.

You pour the batter in and flip the iron like you’re arming a Cold War missile silo.

And the instructions are always the same.

Step 1: Pour batter.

Step 2: Flip iron.

Step 3: Wait for beep.

But the beep never comes.

Now you’ve got six grown adults standing there staring at this waffle maker like it’s a mysterious alien artifact.

Nobody wants to open it too early.

One lady whispers, “I think it’s still cooking.”

How do you know?? Are you a waffle clairvoyant??

Meanwhile there’s always one maniac who just opens it anyway and rips the waffle out like he’s harvesting tree bark from a haunted forest.

Half the waffle sticks to the iron. The other half looks like a burnt map of Italy.

And he’s standing there smiling like he just solved quantum mechanics.

Then you finally sit down with your waffle fragments and coffee that tastes like the color brown, and you look around the room.

Twenty-seven adults.

Same gray hoodies.

Same phones.

Same breakfast strategy.

And that’s when you open your phone and see the Gizmodo article.

“Researchers say AI is homogenizing human expression and thought.”

And I’m reading this thing thinking, THIS ARTICLE is the most homogenized thing I’ve read all week.

Every tech article now sounds exactly the same.

“Researchers warn…”

“Experts say…”

“Concerns are growing…”

Who are these experts? Where do they live? Are they hiding inside the waffle machine?

The structure never changes.

First paragraph: mild panic.

Second paragraph: quote from a professor whose job title sounds like a Wi-Fi password.

Third paragraph: vague cultural implications.

Fourth paragraph: “However, some experts disagree.”

Of course they do. You gotta include the balanced oatmeal paragraph so nobody gets mad.

And then the ending.

Always the same ending.

“More research is needed.”

Oh REALLY?

More research is needed?

You just wrote eight hundred words about it. What were you doing the whole time, practicing typing??

The language in these things always reads like a press release that got a graduate degree in cautious disappointment.

Nobody ever just says the obvious sentence.

“The robot writes like a polite middle manager.”

There! That’s the line! Print that!

But no, it has to sound like:

“Large language models may contribute to convergent linguistic patterns among users.”

What does that even mean?!

That sounds like a weather forecast for grammar fog.

And the funniest part is the article itself is an exact demonstration of the thing it’s warning about.

It’s like writing an article titled The Dangers of Beige Paint entirely in beige paint.

The tone, the structure, the phrasing — all identical to the last twenty-seven tech anxiety articles published this month.

You can practically hear the editorial meeting.

“Okay people, we need a culture piece.”

“Do we have new information?”

“No.”

“Perfect. Let’s say researchers are concerned.”

“Add a professor quote.”

“Ending?”

“More research is needed.”

Boom.

Ship it.

Meanwhile the actual world is still out there being gloriously ridiculous.

Some guy in Finland is making techno music with microwaves and a bicycle wheel.

A kid in Brazil is rapping in three languages while wearing a shark costume.

There’s a grandmother somewhere on TikTok teaching dumpling recipes while arguing with a goose that clearly hates her.

Human weirdness is basically cockroaches. You cannot get rid of it.

But the people writing these articles?

They’ve somehow achieved a level of stylistic uniformity that would make airport carpeting jealous.

And they’re typing this with a straight face saying AI might destroy diversity of expression.

Buddy.

You just wrote the 27th identical article about AI anxiety this month.

If homogenization is the problem, maybe start by looking in the mirror.

Or better yet, come down to this hotel buffet with me.

Stand in line with the rest of us while we stare at the waffle machine like it’s the Rosetta Stone of breakfast engineering.

Because if AI ever truly flattens humanity into one smooth beige cultural pancake…

it’s still going to have to compete with the unstoppable, soul-crushing uniformity of tech journalism.

There is a fairy tale making the rounds, all powdered wig and trembling lip, that music was once a pure human utterance, some sacred exhalation of the species, and then one day AI arrived in a silicon cape and poisoned the village well. This is nonsense. Antique nonsense. Premium nonsense. Music production has been crawling with non-human mediation, prosthetics, substitutions, illusions, ghost labor, mechanical cheats, engineered spaces, synthetic stand-ins, electrical disguises, mathematical corrections, and glorified fraudulence for more than a century.

AI did not introduce the inhuman into music. It wandered into a neighborhood that already had twenty-seven pawn shops, six body shops, and a back alley full of fake mustaches.

And the funny thing is that the people weeping over “soulless” AI will happily listen to a kick drum that is not a drum, in a room that never existed, played in time no person could actually maintain, pitched to a scale no singer ever really hit, written by three writers nobody knows, quantized by a stranger, comped from forty-seven takes, sidechained by software, tuned by algorithm, widened by phase trickery, and blasted through a mastering chain that could flatten a moose.

How about we make a list? Let’s drag every respectable little convenience into the street and make it state its name.

A long list of inhuman elements in music production

1. Drum machines

2. Step sequencers

3. MIDI sequencing

4. Piano roll editing

5. Quantization

6. Groove quantization templates

7. Humanize functions

8. Swing percentage knobs

9. Metronomes

10. Click tracks

11. Looping

12. Sample replacement

13. Drum triggering

14. Beat slicing

15. Time stretching

16. Pitch shifting

17. Auto-Tune

18. Melodyne

19. Vocal comping

20. Take comping for instruments

21. Elastic audio

22. Warp markers

23. Flex Time

24. Flex Pitch

25. Reverb

26. Convolution reverb

27. Algorithmic reverb

28. Delay

29. Slapback delay

30. Tape delay emulation

31. Chorus

32. Flanger

33. Phaser

34. Tremolo

35. Ring modulation

36. Distortion

37. Saturation

38. Tape saturation

39. Console emulation

40. Exciters

41. De-essers

42. Compressors

43. Multiband compressors

44. Limiters

45. Expanders

46. Noise gates

47. Transient shapers

48. EQ

49. Dynamic EQ

50. Match EQ

51. Stereo widening

52. Mid-side processing

53. Sidechain compression

54. Ducking

55. Noise reduction

56. Spectral repair

57. De-click tools

58. De-hum tools

59. De-clip tools

60. Vocal alignment software

61. Drum alignment software

62. Timing correction across multitracks

63. Tuning to equal temperament

64. Tuning to reference pitches nobody in nature requested

65. Layered vocal doubles

66. Artificial doubles

67. ADT, artificial double tracking

68. Harmonizers

69. Octavers

70. Vocoders

71. Talk boxes

72. Formant shifting

73. Granular synthesis

74. FM synthesis

75. Additive synthesis

76. Subtractive synthesis

77. Wavetable synthesis

78. Physical modeling synthesis

79. Sample-based orchestras

80. Synth brass

81. Synth strings

82. Synth choirs

83. Fake pianos

84. Fake organs

85. Fake drums

86. Fake basses

87. Fake handclaps

88. Fake audience noise

89. Foley crowd sweetening

90. Applause tracks

91. Canned laughter in music TV recordings

92. Triggered live backing tracks

93. Playback rigs on stage

94. Stems flown in live

95. Lip-syncing

96. Partial lip-syncing

97. Pre-recorded “live” harmonies

98. Invisible ghost musicians offstage

99. Session musicians who never get public credit

100. Song doctors

101. Topliners

102. Ghostwriters

103. Anonymous arrangers

104. Anonymous orchestrators

105. Copyists who silently shape the playable result

106. Engineers who determine the actual sound more than the performer does

107. Mix engineers who reinvent a song after recording

108. Mastering engineers who alter tone, punch, width, and loudness

109. Producers who dictate phrasing, structure, instrumentation, and feel

110. A&R executives forcing song structure changes

111. Market-tested chorus rewrites

112. Demo singers replaced by stars

113. Session drummers replacing the band drummer

114. Session bassists tightening bad originals

115. Editing out human breaths

116. Editing in human breaths

117. Fake room tone

118. Room mics blended from a different session

119. Re-amping guitars after the performance

120. Amp simulators

121. Cabinet impulse responses

122. Drum sample libraries recorded in rooms you were never in

123. Orchestra libraries recorded by people you will never meet

124. Choir libraries simulating entire languages

125. Phrase libraries

126. Arpeggiators

127. Auto-accompaniment

128. Style engines in arranger keyboards

129. Band-in-a-Box style generation

130. Accompaniment loops

131. Construction kits

132. Royalty-free sample packs

133. Splice packs

134. Sample interpolation

135. Resampling

136. Chopping old records into new songs

137. Tape splicing

138. Varispeed recording

139. Reverse tape

140. Tape flanging

141. Punch-ins

142. Punch-outs

143. Drop-ins for single words or notes

144. Comped “one performance” built from dozens of passes

145. Multi-miking one source to create impossible presence

146. Close-miking that no listener would ever experience in life

147. Isolation booths

148. Gobos

149. DI bass with no speaker in the room

150. Silent recording with virtual amps

151. Grid editing

152. Visual waveform editing

153. Spectral editing

154. Tempo maps

155. Marker-based restructuring after the fact

156. Copy-paste choruses

157. Copy-paste drum fills

158. Cloned backing vocals

159. Synthesized sub-bass under acoustic bass

160. Layering a real snare with four fake snares

161. Layering real strings with synthetic sheen

162. Replacing kick transient with a sample click

163. Pitching vocals note-by-note

164. Level automation making a weak singer sound dynamically masterful

165. Automation rides on every phrase

166. Pan automation

167. Filter sweeps drawn with a mouse

168. LFOs modulating sound with machine regularity

169. Euclidean sequencers

170. Probability-based sequencers

171. Randomization functions

172. Generative modular patches

173. Algorithmic composition tools before “AI” was the buzzword

174. Markov-chain melody generators

175. Rule-based harmonizers

176. Auto-arrangers

177. Chord suggestion tools

178. Key detection tools

179. Beat detection tools

180. Loudness normalization for streaming

181. Brickwall limiting

182. Loudness war mastering

183. Vinyl simulation

184. Tape hiss added for “warmth”

185. Artificial wow and flutter

186. Bitcrushing

187. Sample-rate reduction

188. Stereo imagers

189. Bass mono-izing

190. Master bus glue compression

191. Stem mastering

192. Mastering assistants

193. Mix assistants

194. Smart EQs

195. Automatic gain leveling

196. Automatic de-noise

197. Automatic dereverb

198. AI-powered source separation used by people who hate AI in principle until it saves their hi-hat

199. Drum machine “human feel” presets

200. “Live room” plugins simulating architecture

201. Simulated microphone models

202. Simulated preamp coloration

203. Simulated tape machines

204. Simulated tube warmth

205. Simulated broken speakers

206. Simulated cassette decks

207. Simulated vinyl crackle

208. Simulated radio transmission

209. Simulated distance

210. Simulated intimacy

211. Fake stereo from mono sources

212. Mono compatibility checks shaping stereo decisions

213. Headphone crossfeed compensation

214. Loudness meters policing art by platform standards

215. Broadcast processors

216. Auto-ducking voiceover music beds

217. Jingles built to mnemonic formula

218. Film scoring temp tracks steering composition

219. Trailer braaams nobody in nature has ever heard

220. Risers

221. Downlifters

222. Impacts

223. Sub drops

224. Cinematic booms

225. Braided hybrid orchestral-electronic percussion

226. Trailer choir patches yelling nonsense Latin

227. “Epic” percussion built from dumpsters and processed thunder

228. Choirs tuned and time-aligned like laboratory mice

229. Fake live albums with studio overdubs

230. Crowd noise added in post

231. Audience mics swapped from another night

232. Entire concerts repaired syllable by syllable

233. Tuning forks replaced by digital tuners

234. Strobe tuners

235. Autocorrecting intonation displays for performers

236. Beat detective tools

237. Notation software playback that shapes composing decisions

238. MIDI mockups that decide whether real players ever get hired

239. Temp vocals later replaced

240. “Scratch” guitar parts later kept because the fake became the truth

241. Remote collaboration across files instead of shared physical performance

242. Dropbox and email as ensemble space

243. DAWs as compositional environments

244. Screens as primary musical interface

245. Mouse-drawn crescendos

246. Velocity editing

247. Note nudging by milliseconds

248. Mic tuning by phase alignment

249. Phase rotation tools

250. Latency compensation

251. Offline bounce

252. Stem printing

253. Recallable sessions

254. Total automation snapshots

255. Presets made by strangers determining aesthetic choices

256. Signature producer templates

257. Factory drum kits

258. Genre presets

259. “Make it pop” mastering chains

260. Streaming reference matching

261. Sampled applause for demos

262. Fake count-ins

263. Artificial tape stops

264. Reverse reverbs

265. Infinite sustain plugins

266. Freeze functions

267. One-button chord pads

268. Scale-locking MIDI tools

269. Strum emulation

270. Virtual guitar articulations

271. Breath-noise knobs on fake woodwinds

272. Bow-change scripting on fake strings

273. Legato scripting in sample libraries

274. Round robin samples to imitate human variation

275. Random velocity scripts

276. Random timing jitter scripts

277. Swing extracted from old recordings and applied like digital taxidermy

278. Machine-learned stem extraction from archival recordings

279. Restoration tools reconstructing missing frequencies

280. Spatial audio upmixers inventing space

281. Dolby Atmos object placement beyond real performance experience

282. Binaural rendering

283. Head-tracked audio illusions

284. Psychoacoustic bass enhancement

285. Harmonic synthesis to imply missing fundamentals

286. Auto-mastering web services

287. Algorithmic playlist placement shaping songwriting

288. Skip-rate analytics affecting intros

289. Streaming payout economics affecting song length

290. Hook placement optimized for platform behavior

291. A/B tested chorus entries

292. Intro truncation for social clips

293. Vocal snippets designed for meme velocity rather than performance reality

294. Music supervisors dictating emotional contour

295. Library cues built to metadata categories

296. Mood tags steering composition

297. Corporate “uplifting acoustic” templates

298. “Authentic indie” presets mass-produced by the thousands

299. “Lo-fi” imperfection sold as a controlled product

300. Entire fake nostalgia as a production strategy

That’s three hundred. But there are more. There are always more. Music production is a junk drawer in a mad scientist’s kitchen. Pull it open and there’s a kazoo, a limiter, a payroll sheet, a Mellotron flute, three anonymous Scandinavians, and a plugin called Vintage Air Cemetery.

The central conceit

Notice something. Almost every one of these tools exists for one of a few reasons:

To replace a missing human.

To improve an inadequate human.

To simulate a human.

To organize many humans.

To hide what the humans really did.

To fake the conditions under which humans supposedly made the thing.

Or to produce something no human body, room, ensemble, lung, hand, bow, throat, or memory could have achieved unaided.

That is the history. Not a meadow. A machine shop.

And now, AI shows up

Okay, here comes AI, and soon some people discover that it actually does some alarming little human-ish things. It can drift in tempo. It can bend intonation in expressive ways. It can shape phrasing with a hint of stress, fatigue, tenderness, strain, release. It can do that tiny push into a note that says, without words, “this matters.” It can decelerate not because a grid says so, but because the phrase feels like it has reached the edge of the table and is looking over. Well, not “because,” but “as a result of” having encountered such human messiness in its training.

And what happens next?

Music Industry folks say: lovely, now iron that out.

Not because humans are monsters. Because workflows are monsters.

Here’s what happening: The second these systems began generating output with too much rubato, too much timing elasticity, too much intonational ambiguity, too much expressive contour, too much squishy human unpredictability, the same production culture that has spent decades flattening drummers, tuning singers, snapping bassists to the lattice, normalizing loudness, and replacing rooms with plugins did what it always does. It reached for the wrench.

Make the stems align.

Make the BPM stable.

Make the key center legible.

Make the transients editable.

Make the phrases loopable.

Make the output predictable.

Make it easy to comp.

Make it fit the session.

Make it sit in the mix.

Make it obey.

What are AI music toolmakers like Suno.com doing? AI-generated recordings sometimes contained microtiming and pitch variance—exactly the kinds of “human noisiness” that real musicians use to create feel. And those variances made the outputs hard to import into rigid, grid-centered DAW1 sessions without drift and mismatch.

Their “fix” wasn’t to redesign the pipelines to accept expressive time. The fix was to flatten the expressiveness.

Suno encountered the reality that users—especially professional producers—wanted outputs that could be re-ingested into the assembly pipeline. They wanted stems. They wanted grid-aligned timing. They wanted predictable BPM. So they introduced workflows to make that possible, including tools that smooth timing and help integrate with legacy systems.

Then came the more recent shift people are now experiencing: newer models like Google’s Lyria3 and ElevenLabs Music that sound increasingly “professional” in the narrow sense—more prominent vocals, smoother timing, notes pulled toward equal-tempered pitch.

If you’ve compared outputs across these tools, you can hear it: with each new update or release, the performances get tidier. More “correct.” Less risky. More like what the system already rewards.

So the tools are becoming not just generators of music, but enforcers of a particular industrial template.

That is the delicious insult in all this. The supposedly terrifying machine became truly useful only when it was disciplined into the older machinery. The parts of AI that most resembled living musical behavior were quickly treated like a drummer who came in with “feel.” Everybody nodded solemnly, then dragged him into post and beat him with a quantizer.

The machine was allowed to be “creative” right up to the point where creativity got in the way of production.

You can practically hear the meeting:

“Can AI make it more human?”

“Yes.”

“Umm. Can it make it less human so we can ship it by Thursday?”

And there it is. Civilization.

But what about the artists?

Subscribe now

The story that music was pure, human, and unsullied until AI came along is wrong for several reasons.

First, music has never been just a naked human expression pouring directly from soul to ear. Music has always been mediated by tools, training, rituals, conventions, instruments, notation, acoustics, commerce, technology, institutions, patronage, labor hierarchies, and fraud. A cathedral organ is not less artificial because it is old. A violin is a machine. Equal temperament is a compromise racket. Conducting is social engineering with cufflinks.

Second, what people call “human” in music is usually not raw humanity at all. It is familiar stylization. It is a culturally approved package of deviations, imperfections, balances, expectations, and signs. The right kind of roughness. The acceptable wobble. The curated crack in the voice. The sanctioned drag behind the beat. The approved amount of blood on the shirt.

Third, people are sentimental liars about the past because sentimentality is one of the few luxuries the nervous system can afford. It is totally human to look backward and imagine that once, before the latest horror, things were cleaner. Before multitrack. Before synths. Before drum machines. Before Auto-Tune. Before samplers. Before electric guitars, for that matter, which sounded to some earlier bores like the very growl of the apocalypse.

Every generation appoints itself guardian of the last honest tool. Then that tool gets absorbed into the wallpaper and the next thing becomes the devil. This may be one of humanity’s great hobbies, somewhere between gossip and arson.

The hidden truth beneath the panic

Here is the richer, more embarrassing truth:

All this “non-human” stuff is something only humans do.

Only humans build a machine to imitate another machine that imitates a horn section that was itself arranged to sound like a cartoon version of a military bugle filtered through three decades of nightclub smoke.

Only humans invent a plugin to make a sound less real, then another plugin to make the fake sound more vintage, then a third plugin to make the vintage fake sound emotionally trustworthy.

Only humans say they want authenticity and then spend fourteen hours removing saliva noises, pitch variance, room bleed, tempo drift, chair squeaks, and every other microscopic trace that an actual organism was present.

Only humans hire ghostwriters and then pontificate about authorship.

Only humans create a workflow so sterilized that they later need a “humanize” button to put back the crumbs of life they vacuumed up themselves.

Only humans can be so hungry for control that they mechanize feeling, and then so hungry for feeling that they buy simulations of lost mechanisms.

So the inhuman in music was never the opposite of the human. It was one of humanity’s favorite pastimes. It is a human compulsion to externalize, delegate, disguise, automate, idealize, and re-import expression through contraptions. We do it in law, in war, in architecture, in religion, in dating, in cuisine, in government, and yes, in music. Especially in music, because music gives us the flattering illusion that our manipulations are noble.

What the AI music panic is really about

It’s not really about soul. Soul is the decorative word people reach for when they do not want to describe labor arrangements, authorship anxiety, market destabilization, status threat, expertise dilution, and the terrifying possibility that the line between “natural gift” and “assembled result” was always blurrier than the priesthood admitted.

That is what rattles people.

Not that AI is inhuman. That it reveals how industrial, assisted, composited, outsourced, and synthetic music already was before AI.

It throws a hideous fluorescent light on the room. Suddenly the old compromises are visible. Suddenly the magic trick has wires. Suddenly the singer is a committee, the room is a plugin, the drummer is a spreadsheet, the orchestra is a hard drive, and the authenticity merchant is standing there with twelve layers of vocal tuning explaining that this new thing is fake.

Of course people hate that. It’s a rotten feeling when the latest technology does not merely threaten your livelihood but insults your mythology. Especially if it can produce artifacts that actually fit the constraints we’ve been accreting for 70 years.

Do not be stupid.

This does not mean all criticism of AI music is wrong. Some of it is correct. Plenty of AI output, like human output, is bland, derivative, over-smoothed, statistically obedient paste. Plenty of it is designed to exploit rather than illuminate. Plenty of companies treat artists like disposable training mulch. Plenty of users prefer speed over thought. There are serious questions about consent, compensation, attribution, cultural theft, and industrial scale pollution.

Fine. Good. Make those arguments.

But spare us the church-lady act about purity.

Do not say that the sanctity of music was intact until the chatbot in a leather jacket showed up. Music has been a glorious racket, a relay race of masks and machines, for ages. The problem with AI is not that it introduces artifice into an Eden of sincerity. The problem is that it arrives inside an already mechanized culture and can be used either to deepen that machinery or, in some rare cases, to expose it.

So what does it all mean?

It means that humanity is not the absence of tools. Humanity is the animal that cannot stop making them.

It means that “real” in music has never meant “untouched by mediating technology.” It usually means “mediated in ways I am accustomed to, by classes of people I have chosen to respect, with technologies old enough that I no longer notice them.”

It means the boundary between human and non-human in art is not a clean border but a filthy estuary, full of silt, cables, anonymous labor, mythology, aspiration, fear, and monkey-level status games in expensive eyewear.

It means that what we love in music may not be purity at all, but the successful organization of artifice into an experience that feels alive.

And it means that the proper response to AI is not infantile nostalgia, but adult discrimination. What is this tool doing? Who benefits? Who is erased? What forms of expression does it flatten? Which ones does it unlock? Where does it let laziness metastasize? Where does it let imagination run like a lunatic through the garden?

That is the real conversation.

Not this powdered corpse of an argument that once there were Real People making Real Music in some glowing pasture, until silicon came clanking over the hill.

But isn’t AI still bad?

Most people reading this are not crouched over a DAW at 2:30 in the morning arguing with a compressor about whether the snare drum is “punchy but warm.” Most people are not auditioning microphone models or wondering if the ride cymbal needs three milliseconds of negative delay to sit in the groove.

Most people just… listen.

They hear a song in the car. They hear something in a café. A melody drifts out of a phone speaker while someone is cooking. The bass feels good. The chorus hits. The thing works. End of transaction.

And that is perfectly reasonable. Nobody expects someone eating boneless chicken wings to perform a forensic autopsy on poultry logistics.

You order the wings. They arrive. You eat them. They taste good.

You do not pause halfway through to ask whether the bones were removed by surgeons with tweezers, or whether chicken slurry in a vat the size of a city bus was extruded through a wing-shaped mold. You do not interrogate the structural integrity of the breading matrix. You do not demand a white paper on the metaphysics of poultry.

You eat the wing.

Music works the same way.

We don’t sit there thinking, “Ah yes, the second verse vocal was probably comped from fourteen takes, and I suspect the bass was replaced by a DI through a Neve emulation with parallel compression.” We hear the feeling of the thing. That’s it.

For most of the last hundred years, the bones of music production have been quietly removed in the kitchen. Carefully, methodically, invisibly.

None of this was done maliciously. It was done because humans are tireless tinkerers. We cannot leave anything alone. If there is a stick nearby, we will hit something with it. If there is a string, we will tighten it. If there is a hollow bone lying around, someone will eventually blow through it.

And that brings us back about sixty-thousand years.

Gronk started it all

Somewhere in the Ice Age, some bored, curious primate named Gronk, perhaps, picked up a piece of bear bone. Maybe it already had a few holes in it. Maybe the holes came later. Nobody knows. But eventually Gronk lifted it to their mouth and blew through it.

Out came a tone.

Not a perfect tone. Not a conservatory tone. Just a sound.

And that was enough.

That moment did not end music’s purity. It began music’s long career as a technological experiment.

Because that bone flute was already cheating.

The voice alone could only do so much. The bone extended it. It amplified it. It stabilized the pitch. It let a person hold a tone longer than lungs would normally allow. It created a new color that did not exist in the naked human throat.

It was a tool. An instrument. And once humans start using tools, the rest of the story practically writes itself.

• The stick becomes a drumstick.

• The drumstick meets a stretched skin.

• The skin becomes a drum.

• The drum becomes a drum set.

• The drum set meets microphones.

• The microphones meet tape.

• Tape meets editing.

• Editing meets multitracking.

• Multitracking meets synthesizers.

• Synthesizers meet computers.

• Computers meet software.

• Software meets algorithms.

And eventually, algorithms meet something that looks suspiciously like intelligence.

At each step along the way, somebody declared the new thing a scandal. And now AI is the latest guest at the outrage banquet.

But if you zoom out just a little, the pattern becomes impossible to miss. Humans pick up whatever is lying around and make music with it. The instinct itself has never changed. The tools just keep getting weirder.

What will we do with AI?

History suggests the answer will not be boring.

When tape arrived, people did not merely record concerts. They invented studio music that could not exist on a stage.

When synthesizers arrived, people did not merely imitate horns. They invented sounds nobody had heard before.

When samplers arrived, people did not merely copy old records. They also built entirely new rhythmic languages out of fragments.

Technology does not replace musical imagination. It rearranges the playground.

Yes, some people use AI to produce endless wallpaper music. That will happen. It always does. Every new tool first produces a tidal wave of mediocre convenience. That mimetic phase is inevitable. It’s how we learn.

But somewhere inside the same technology, somebody is going to do something astonishing with it. Someone will bend it. Misuse it. Combine it with a violin, a choir, a modular synth, a field recording of rain, and a strange rhythmic idea that arrived during a bus ride.

And suddenly the thing will feel new.

That is how music moves forward. Not by preserving purity, but by experimenting with impurity.

Who’s really worried now?

Of course, a tiny but vocal slice of the current music economy is understandably nervous. There are business models tied to the old production pipeline. There are professions that depend on the old bottlenecks. There are workflows built around scarcity.

Those concerns are real for the people whose livelihoods depend on them.

But for the listener, the ordinary civilian who simply wants to hear something beautiful or strange or moving on a Tuesday afternoon, the situation is much simpler.

More tools mean more music.

More strange combinations.

More unexpected sounds.

More people who previously had no way into the process, now suddenly able to participate.

Some of it will be awful. Some of it will be trivial. Some of it will be the musical equivalent of a boneless wing assembled in a factory of questionable ethics. But somewhere in the pile will be something wonderful.

There always is.

So maybe the healthiest reaction to AI in music is not purity panic. Not hand-wringing over the death of the soul. Maybe the better response is a different kind of hand-wringing. The anticipatory kind. The kind where you lean forward a little and wonder what the next generation of odd, brilliant, stubborn humans is going to do with the new toys scattered on the floor.

Because if sixty-thousand years of evidence tells us anything, it is this: Give humans a hollow bone, a strange machine, or a piece of software nobody fully understands…

…and before long, someone will make music with it.

And when they do, we will not ask how the bones were removed.

We will just listen.

Copyright © 2026 by Paul Henry Smith

I want you to picture a very specific person, because this is not a philosophical debate. This is a career situation.

You’re a UX designer. You love Figma. You love the feeling of turning a messy problem into clean, tasteful UI. You love speed. You love craft. You love being the person who can crank out a polished flow while everyone else is still arguing about what the feature even is. Your portfolio is screens, screens, screens—beautiful, consistent, modern screens—and hiring managers love you for it. They barely read anything. They scroll. They nod. They go, “Yep. This person can ship.”

That person might be you. It might be your teammate. It might be half the industry.

Now here’s the moment that made me stop and stare at the wall for a while: I saw a job posting from PayPal that wasn’t shy about where this is going. It wasn’t “AI-assisted design tooling.” It wasn’t “copilot for designers.” It was basically: we want to automate the production of UI and connect it directly to live business inputs—revenue, conversion, telemetry, trends, real-time analysis, prediction—and then generate solutions continuously.

In plain English: the system sees a signal and changes the interface. Constantly. All day and all night.

And if PayPal is willing to say that in public—if they’re comfortable putting that vision in a job description—then you should assume everybody else is thinking it too, even if they’re being quieter about it. Because nobody wants to be the second company on earth to admit they’re trying to automate a whole profession. They want to be the first company to quietly succeed and act like it was “obvious.”

So if you’re sitting there thinking, “Yeah, but they can’t replace me, I have taste,” I need you to understand something, and I’m going to say it bluntly because it’s kinder than letting you keep believing it:

Subscribe now

Taste is not a moat when your taste has already been turned into rules.

Most modern design teams spent the last decade doing something that was genuinely smart: standardizing. Tokens, components, pattern libraries, accessibility rules, spacing systems, interaction conventions. It made teams faster. It made products more consistent. It reduced chaos.

But it also did something else—something we didn’t want to think about because it ruins the vibe.

It made the work legible.

If your product has a design system that dictates what “good” looks like, then a lot of downstream UI design becomes: pick the correct component, apply the correct pattern, follow the rules, don’t break anything.

That’s not an insult. That’s how you scale.

But it also means the work is learnable in the way machines love: lots of examples, lots of constraints, lots of “approved vs rejected,” lots of history.

You don’t need a machine that understands beauty. You need a machine that predicts what will pass design review.

And we have built an entire industry around making that prediction easier.

Now, before you get mad at me, let me be fair to everyone involved, including the so-called “Figma farmers.”

A lot of designers didn’t choose to be trapped in UI polish. They were trained into it. They were hired into it. They were rewarded for it. They got promoted for it. And they got hired in the first place because that’s what our hiring processes selected for.

This part matters, and it’s not comfortable: during the pandemic hiring boom—when everyone was hiring like drunk sailors—UX teams didn’t scale by carefully selecting for deep systems thinking. They scaled by selecting for what could be evaluated quickly.

Screens.

We did it. I did it. I sat in interview loops. I watched people scroll portfolios like they were browsing Zillow. “Look at the craft.” “Look at the polish.” “Look at the number of flows.” “Look how fast they can produce.”

And bootcamps, being rational businesses, trained people to win that game. They didn’t train “how to kill a feature with a principled argument.” They trained “how to present a case study with a gorgeous Figma flow.” Because that’s what got interviews.

So it’s not that product and engineering forced design into a corner and design heroically endured. The uglier truth is that design, under pressure and incentives, overselected for visible output. We trained ourselves to prove our worth with artifacts.

And now the artifact factory is being automated.

That’s the part that should piss you off—not at the designers, but at the incentive structure we all participated in, because it’s about to cash out.

Now let’s get to the real heart of it, because if this were just “AI makes pretty UI,” it’d be annoying but manageable.

The real thing PayPal is going after is latency.

Traditional UX is slow in a very specific way. Not because designers are slow. Not because teams are dumb. Because the loop is human.

A metric moves. Someone notices. Someone convinces others it matters. Research happens. A fix is designed. It gets reviewed. It gets built. It ships. The world changes again.

PayPal’s vision is: skip the human noticing-and-coordinating part. Wire the system directly into the signals. Let it propose and implement UI changes continuously.

That is a very different world. In that world, “being fast in Figma” is not a flex. It’s like bragging that you’re the fastest person alive at hand-washing dishes while the restaurant installs an industrial dishwasher.

You’re standing there, sleeves rolled up, like, “Guys, watch me go!” and management is like, “Yeah… cool… anyway…”

Now, this is where people either get defensive or go numb, so let me ground it in two very real anecdotes, because otherwise this stays abstract and you can keep comforting yourself.

When I worked at AWS, an old friend called me and said—exact words—“Why does setting up IAM roles make me want to commit murder?”

Now, was he being dramatic? Sure. But he wasn’t wrong. IAM is not confusing because the buttons are ugly. It’s confusing because the system’s mental model—the way it thinks about identity, permissions, relationships—does not map cleanly onto how humans think about responsibility and access. It’s architecture-first. You, the user, are being asked to understand the machine’s view of the world and behave accordingly.

And what does the organization do with that? It doesn’t say, “Let’s rethink how permissions should be modeled for human beings.” It says, “UX, make it clearer.”

Which often means: build explanatory UI around an unchanging architecture. Make it navigable. Make it survivable.

The second story: I did research with music students using SmartMusic. Teenagers, yes, moody, yes—but listen to what they said. “SmartMusic makes me want to slit my wrists.” And then, quieter but worse: “SmartMusic is what made my brother quit music.”

That isn’t about UI polish. That’s about the interaction contract: what the system demands from you, what it remembers, what it punishes, what it rewards, and how it makes you feel while you’re trying to learn. It’s cognitive and emotional architecture. The interface is just the messenger.

These two stories are extreme versions of something every designer has seen: the “UX problem” is often that the product was built around the system’s mental model, not the user’s. The UI is then asked to translate the system’s worldview into something humans can tolerate.

That’s explanatory UI. And yes, it is real work. It’s hard work. It takes skill. It is not trivial. But it is exactly the kind of work that an AI system—given enough examples—can start doing at scale, because it lives downstream of decisions that are already made.

And here’s the crucial point, which is the one that actually matters if we’re talking about a future where machines generate interface all day: the real design work is not just “should we build it or not.” It’s designing the dance between system and person.

It’s deciding what the system should know and remember, and what the human should know and remember.

That sounds subtle until you realize it’s basically the whole game.

If your product makes the user remember fifteen things the system could easily remember, you’re building stress. If your product hides state the user needs to understand, you’re building confusion. If your product demands the user maintain the system’s internal picture of the world in their head, you’re building anger. If your product pushes critical memory into tooltips and docs and “learn more,” you’re building failure.

Design at its best is a kind of cognitive engineering. It’s deciding where the burden goes, and making the burden land where humans are actually good at carrying it. Humans are good at recognizing patterns, forming habits, and navigating a consistent mental model. Humans are bad at holding lots of arbitrary state, tracking invisible rules, and recovering from unclear errors without feedback.

And here’s the problem: most organizations accidentally design products that require humans to do exactly what humans are bad at, because that’s what the architecture made easiest.

Then they hire a designer to paint over it.

Now let’s talk about the scary “third outcome”, because it’s not just PayPal building a continuous optimization machine. It’s everybody vibe-coding like lunatics. PMs, engineers, VPs—all of them generating screens, flows, features, filters, settings, clever little options. The tools are so helpful they can’t stop themselves. And humans, being easily seduced by possibility, keep saying, “Sure, add that too.”

In that world, UI becomes an all-you-can-eat buffet run by a robot chef who never sleeps and never gets tired of adding “one more thing.”

Your careful, tasteful Figma work doesn’t look valuable. It looks slow. It looks fussy. It looks like you’re polishing a spoon while the kitchen is flooding with pasta.

So the move is not “I will generate more UI too.” That’s suicide. You cannot win the output contest. Output will become infinite.

The move is to become the person who can look at infinite output and say, calmly and clearly, “Most of this is noise, and here is why.”

Not as an aesthetic judgment. As an interaction judgment. As a mental model judgment.

This is where I need to be skeptical, too, because I don’t want to turn PayPal into the boogeyman and pretend they’re guaranteed to succeed. They might not. Systems wired to business metrics tend to optimize what’s measurable, and what’s measurable is often short-term. They can overfit. They can drift. They can create interfaces that juice conversion while eroding trust, comprehension, or long-term satisfaction. They can make products feel like slot machines—always adjusting, always nudging, never letting the user build a stable understanding of what the system is and how it behaves.

And when that happens—when users start saying “I don’t trust this thing,” or “it keeps changing,” or “it feels manipulative,” or “I can’t predict what it will do”—that won’t show up cleanly in your dashboard until it’s already a problem.

Which means there is still human work here. But it’s not the work most Figma-first designers have been trained for, and that’s the part we have to say without being cruel about it.

If you’ve been rewarded for speed and polish, it doesn’t mean you’re dumb. It means you played the game in front of you. If you got hired because your portfolio showed pixel-perfect flows, that’s not because you’re shallow. It’s because that’s what the market selected for. We did that. Our teams did that. Our interview loops did that. Our bootcamps responded to that. The whole ecosystem reinforced it.

But now the ecosystem is changing, and the old signal of competence—screens—won’t mean what it used to mean.

So what do you do, practically, if you’re that gung-ho designer who loves craft and doesn’t want to become an “AI policy person,” and also doesn’t want to be automated?

You don’t have to become a philosopher. You don’t have to become a PM. You don’t have to become “strategic” in a buzzword way. You have to move one layer earlier than the screen. You have to start designing interaction contracts: what the system remembers, what the user remembers, what feedback is given when things go wrong, what state is visible, what is hidden, and why.

You have to start caring about architecture—not the backend details, but the human-facing shape of it. What is this thing? What does it believe about the world? What does it require the user to believe? Can a normal person form an accurate mental model without needing a wiki?

And yes, you can start doing this even if your org is messy. You can do it in small slices. You can do it in the way you frame problems. You can do it by asking better questions before you produce UI. You can do it by writing clearer system stories: “Here’s what the system knows at this point; here’s what the user thinks it knows; here’s the gap; here’s where confusion happens.” You can do it by designing for recoverability instead of perfection. You can do it by treating “the user must remember X” as a design smell, not a requirement.

I’m not saying it’s easy. I’m saying it’s possible. Over the last year, as generative AI got serious, I found myself spending more and more time exactly there—because it’s the only place where the work doesn’t collapse into “just generate another screen.” It’s the place where you can still create clarity that isn’t cosmetic. It’s also the place where you can still make a system feel honest, stable, and learnable instead of twitchy and optimized.

Now, I’m not going to end this with some smug “so the future belongs to…” speech. That’s corny. Also, nobody knows. There are too many variables, too many organizational politics, too many ways this could go sideways.

But I do think a few things are likely.

It’s likely that UI production gets cheap enough that “fast in Figma” stops being rare. It’s likely that companies will try to wire optimization loops directly into interfaces, because it’s the obvious move if your goal is moving metrics. It’s likely that some of these systems will work well enough to change hiring immediately. It’s also likely that some will fail in ways that create new kinds of UX disasters—products that are constantly “improving” and yet increasingly incomprehensible.

And it’s likely that the designers who keep their leverage won’t be the ones who can generate the cleanest screens fastest. They’ll be the ones who can make the system-user dance make sense: who can decide what the system should carry, what the user should carry, and how to make that trade visible and humane.

So if you’re reading this as a Figma-loving, craft-proud designer, I’m not here to dunk on you. I’m here to tell you the truth I wish someone had told me earlier: a beautiful explanatory interface is still explanatory. It can still be a mask. And in a world where machines can generate masks all day, your job is not to become a faster mask-maker.

Your job is to stop building things that need masks in the first place, and when that’s not possible, to redesign the interaction contract so the user isn’t forced to carry the system’s architecture in their head like a punishment.

You can be mad at PayPal. You can be mad at AI. You can be mad at the hiring boom and the bootcamps and the “screens-only” portfolio culture. But after you’re done being mad, you still have to do something with that information.

Look, this isn’t about humiliating you for loving craft.

It’s about not letting you shrink your entire identity to a tool.

You’ve got thirty, maybe forty years left in this career. Forty years. If you think you’re going to be lovingly adjusting auto-layout constraints in 2065 like some digital watchmaker, that’s about as likely as someone making a living today hand-coloring black-and-white photographs for the newspaper.

Tools change. Entire mediums change. Nobody sits around crying that they’re not typesetting by hand anymore. The people who survived didn’t cling to the tool. They followed the work.

And the work is not “making screens.”

The work is shaping how humans and systems deal with each other.

That’s bigger than Figma. That’s bigger than AI. That’s bigger than whatever tool gets hot next year.

You’ve already proven you can learn tools. You did it once. You can do it again. That’s not the impressive part.

The impressive part—the part that actually makes you dangerous in a good way—is whether you can look at a system and say, “No. This is the wrong way for a human to have to think about this.”

Whether you can design the memory of a system so people don’t have to carry it around like a backpack full of bricks.

Whether you can see where automation should stop, not because it’s unethical in a hand-wringing way, but because it makes the interaction worse.

That’s not pixel pushing. That’s not vibe coding. That’s not aesthetic judgment.

That’s grown-up design.

And if you lean into that—if you start practicing that instead of obsessing over how fast you can produce variants—you’re not shrinking. You’re leveling up.

Yeah, the machine is coming for the easy stuff. Good. Let it. Why are you fighting to keep the repetitive part of your job? Let it have that. You’ve got better things to do.

You’ve got decades ahead of you. Decades to build things that don’t make people want to commit murder setting up permissions. Decades to build systems that don’t make kids quit music. Decades to shape how automation actually behaves in the real world.

That’s not doom. That’s an opportunity hiding inside an uncomfortable truth.

So don’t walk away from this sulking. Walk away thinking, “Alright. Fine. If the game is changing, then I’m changing with it.”

The world moves. So move with it.

Copyright © 2026 by Paul Henry Smith