I recently stopped using Claude as my daily driver.

I moved most of my work to Codex for a straightforward reason: Claude Fable 5 was too expensive for the way I work. I could give it a serious project, watch it consume my usage limit, and then find myself rationing the model I was paying to use.

The intelligence was not worth much if I could not afford to keep it working.

Claude Fable 5.1 gives me a reason to reconsider that decision.

Anthropic released Fable 5.1 today as its most capable generally available model for coding, research, knowledge work, and long-running agents. The company says it performs similar to or better than Fable 5 at Low or Medium effort for much less money. It also reports major improvements in business-workflow automation, computer use, and multistep research.

Those are exactly the areas that matter to me.

But this is not a review. I have not fire-tested Fable 5.1 enough to tell you whether Anthropic’s claims survive real work. I want to explain why the release has my attention, what I will test, and why another change in the release may ultimately matter more to anyone publishing writing with AI.

Every piece of text produced by Fable 5.1 is now watermarked.

The cost story is the first thing I care about

Fable 5.1 does not have a lower headline token price than Fable 5. Both cost $10 per million input tokens and $50 per million output tokens through the API.

The difference is how much work Anthropic says you can get from those tokens.

Cache reads now cost $0.25 per million tokens instead of $1—a 75% reduction. This matters for long-running agents because they repeatedly reread the same instructions, files, history, and operating context. Anthropic estimates that the change makes a typical Fable 5.1 workload approximately 25% cheaper and a highly agentic workload as much as 45% cheaper.

Effort settings could matter even more in everyday use.

Anthropic says Fable 5.1 at Medium effort roughly matches Fable 5 at a lower cost. At Low effort, it can compete with smaller Claude models on cost per completed task while producing better results in some evaluations.

That could turn Fable 5.1 Low into a strong daily driver. Instead of paying maximum-intelligence prices for every ordinary task, I could reserve the highest effort for decisions that actually require it.

There is an important caveat.

The API economics and Claude subscription limits are not the same thing. Anthropic says Fable 5 and 5.1 currently work the same way on paid plans. Max subscribers can use Fable for up to half of their weekly limit before switching models or using additional credits, and the Fable models consume those limits faster than ordinary Claude models.

So I am not ready to say Fable 5.1 solves the problem that pushed me toward Codex. The honest claim is narrower: Anthropic has given me a credible reason to test whether it can deliver more useful work before I hit the wall.

The benchmark jump that matters to me

Most model benchmark charts do very little for me. I care about whether a model can perform the work sitting in front of my businesses.

Two reported improvements stand out.

Fable 5.1 scored 31.4% on Anthropic’s AutomationBench evaluation for business workflows, compared with 17.1% for Fable 5. That is close to double the score.

Its strict computer-use score increased from 36.1% to 41.7%. Its partial-credit computer-use score rose from 72.9% to 77.9%.

“Computer use” can sound abstract until you operate a real business online.

Substack does not give me a public API for every part of running a publication. Updating publication infrastructure means opening the website, moving through settings, checking the profile, checkout, offers, welcome emails, lifecycle emails, and logged-in and logged-out views. An agent has to work through the same interface a person uses.

The opportunity is not merely an AI that can click buttons. I need an agent that understands the business objective, finds every relevant surface, preserves consistency, avoids changing anything outside the approved scope, and stops when it reaches a decision or publishing gate that belongs to me.

That is what useful computer use looks like.

Deep research is also business infrastructure

I use the highest levels of AI reasoning for work that would otherwise take days.

Recently, I researched six competing creators and businesses before redesigning a strategy. That meant studying their content, products, funnels, free and paid memberships, community models, and revenue structures. The valuable result was not six summaries. It was the synthesis that allowed me to compare the models, reject the wrong assumptions, and decide what fit my business.

The same applies to agency strategy, client acquisition, market research, and planning across several owned ventures.

Anthropic reports that Fable 5.1 more than doubled Fable 5’s score on its scientific terminal-research evaluation, from 24.7% to 52.6%. I do not operate a scientific laboratory, but the underlying behavior matters: searching across many sources, using tools, following a question for a long time, recovering when a path fails, and returning with something coherent enough to inform a decision.

That is the kind of research I want a frontier model to perform.

Thirty-eight hours is impressive. Staying inside the lines is more important.

One of Anthropic’s early users describes an unattended Fable 5.1 run that lasted 38 hours. Another says the model built a complex prototype over approximately three days.

That sounds insane because it is a very different operating model from opening a chatbot, asking a question, and copying the answer.

I run multiple ventures as a solopreneur. Some are full publications with their own audiences, editorial rules, research systems, publishing queues, and automation. One of my owned publications maintains more than 30 assets across an evergreen library and an active news operation. The work includes recurring verification, future review dates, state changes, publishing records, and human editorial gates.

An agent that can work across that system for hours could be enormously valuable.

But duration by itself is not the standard.

The agent has to remember which business it is operating, follow the rules for that publication, preserve the boundary between ventures, maintain its state, verify the work, and stop before taking an action it is not authorized to take.

Running unattended for 38 hours is impressive. Running unattended for 38 hours without losing the plot is what would make it useful.

Can Fable finally write like a normal person?

This is one of my biggest open questions.

I did not enjoy talking to Fable 5, and I did not like much of the writing it produced. It often sounded like I was talking to a PhD who wanted to make sure I knew he was a PhD. The output could become long, dense, overly explanatory, and full of prose that performed intelligence instead of communicating clearly.

That matters if Fable 5.1 is going to produce anything readers actually see.

Anthropic says the new model uses fewer stock phrases and less unexplained jargon. But even its own prompting guide acknowledges that Fable 5.1 can produce denser prose than Fable 5, with longer sentences and fewer paragraph breaks. Anthropic recommends explicitly telling it to remove “mannered prose.”

That does not sound like a solved problem. It sounds like something that needs to be tested with real voice evidence and real publishing standards.

I will be looking for whether Fable 5.1 can:

• follow a distinctive voice without flattening it;

• communicate directly rather than academically;

• stop explaining after the point is clear;

• preserve original observations and arguments;

• avoid stock AI structure and mannered language; and

• produce writing I would willingly publish under my own name.

Until I run those tests, I cannot recommend or reject it as a writing model.

Fable 5.1 text is watermarked

This may become the most important part of the release for anyone using AI to produce public writing.

Fable 5.1 places an invisible statistical watermark in the text it generates. The watermark applies across Claude, Claude Code, Cowork, the API, supported cloud platforms, and every region where the model is offered.

It does not add hidden characters. It does not attach your name, organization, account, or chat history. It does not visibly label a paragraph as AI-generated.

Instead, the model changes the source of randomness it uses when choosing among multiple reasonable next words. Over a sufficiently long passage, those choices form a pattern that someone with Anthropic’s detection key can test.

Anthropic says it introduced the system to comply with the EU AI Act’s transparency requirements. It is applying the watermark globally because it does not yet have a durable way to restrict it by region. Other older Claude models are expected to receive marking support later.

This is not a hypothetical policy announcement. Fable 5.1 is watermarked from launch.

What the watermark proves—and what it does not

A detected watermark means Claude was probably involved in processing a passage.

It does not prove that Claude originated the ideas. It cannot distinguish “Claude wrote this from scratch” from “Claude heavily edited this.” It does not determine ownership, authorship, accuracy, originality, or quality.

It is also not infallible.

Anthropic says detection becomes more reliable as a passage gets longer. Short text may not provide enough signal. Light editing may leave the watermark intact, while heavy editing, paraphrasing, translation, or combining the passage with other writing may make it undetectable. Proofreading human writing may not introduce enough model-chosen language to register at all.

The detection API is currently in private preview for eligible regulators, media organizations, fact-checkers, researchers, educational organizations, civil-society groups, and enterprises with relevant compliance obligations. Your average reader cannot currently paste an article into a public Anthropic detector.

But the direction is clear.

The assumption that AI involvement will remain permanently invisible is beginning to break down.

This makes original substance more important, not less

There is already a frenzy to label writing as “AI slop.” Watermarks will make that conversation more complicated because they can identify likely AI involvement without telling us whether the work is any good.

That distinction matters.

If you give a model no real experience, evidence, observation, argument, or editorial direction, the result is disposable whether it carries a watermark or not.

If you begin with something real, decide what you believe, supply the evidence, direct the process, challenge the output, and accept responsibility for the final piece, the fact that AI touched the text does not erase your contribution.

But you need to be able to explain and defend that contribution.

The response to watermarking should not be a race to disguise AI-generated writing. It should be a better standard for using AI:

• bring original material;

• make the editorial decisions;

• understand what the model contributed;

• review every consequential claim;

• comply with the disclosure rules that apply to your work; and

• stand behind the result.

If a watermark appears, “but it sounds human” is not a meaningful defense. The defense is that the work contains something valuable, you can account for how it was produced, and a responsible person or organization owns the final result.

One feature agent builders should watch

Fable 5.1 also introduces the ability to change effort between messages through the API.

This sounds technical, but the operating idea is simple: a long-running agent no longer needs to use the same intelligence budget for every stage of a project.

It might use High effort to analyze conflicting evidence or set a strategy, Medium effort to perform normal implementation, Low effort for straightforward transformations and administration, and then return to High effort for final verification.

I would not blindly front-load all of the intelligence into planning and run every execution turn at Low. Execution can contain the most consequential decisions, and Anthropic warns that Low effort is less likely to search or retrieve evidence.

The better design is to assign effort according to the decision being made in each stage.

That could make long-running agents much more practical for people who cannot afford to run every turn at the maximum setting.

What I will test next

I am interested in Fable 5.1, but I am not ready to give it a victory lap.

I want to test it against work I already understand:

1. Daily cost and limits. Does Low or Medium effort let me complete meaningfully more work before I hit my plan limits or need usage credits?

2. Business research. Can it reproduce a multistage competitor-research project without skipping verification, flattening differences, or producing a pile of summaries instead of a decision-ready synthesis?

3. Computer use. Can it inspect and prepare changes across a browser-only platform such as Substack without drifting outside the scope or crossing a publishing gate?

4. Long-running operations. Can it maintain state across a real publication workflow, recover from failures, and finish the complete job without mixing contexts?

5. Writing quality. Can it follow a given voice, remove dense and mannered prose, and produce something worth publishing rather than merely something technically correct?

6. Effort routing. Where can I reduce effort without losing the research, verification, or judgment the task requires?

Fable 5.1 could pull me back toward Claude. If Anthropic’s claims survive my actual workloads, it may be one of the first frontier models that gives a small operator enough intelligence, endurance, computer use, and cost control to help run serious businesses every day.

At the same time, its watermark draws a line under the era of assuming AI-assisted writing will always be invisible.

Those are both meaningful changes.

Now they need to survive real work.

—Alex

Founder of WriterOps

It begins with better substance.

This is easy to miss because AI is very good at producing something that looks finished. Give it notes, a topic, an audience, and a desired tone, and it can return a clean draft in seconds.

But a clean draft can still have no real point.

The problem is not always that the user gave AI nothing. Plenty of reasonably experienced users bring voice memos, transcripts, research, meeting notes, examples, and rough ideas into the conversation.

The problem is that they move from raw material to prose before deciding what the material actually means.

That leaves AI to make the central editorial decisions like:

• Which idea is worth building around?

• What does the example prove or reveal?

• Why should this particular reader care?

• What is the piece actually trying to say?

The result may be readable or accurate. But if the model chose the interpretation, the writer no longer has clear ownership of the point.

Share

Context is not the same as a developed idea

Adding more context can improve an AI response, but context alone does not solve this problem.

A folder of research is not a thesis.
A transcript is not an argument.
A personal story is not yet a lesson.

Those are just materials. A piece of content begins when the writer makes an editorial decision about what those materials mean and why they matter.

That is what separates a piece with a point of view from a competent summary of the topic.

The four things you need before drafting

You do not need a finished outline or a perfectly worded thesis before using AI.

Before asking it to draft, however, you need four things.

1. A specific reader or situation

Who is the piece for, and what is happening when they encounter it?

“Business owners” is an audience category. “An expert who has something useful from their work but struggles to turn it into a clear post” is a reader in a situation.

The second gives the piece a job to perform.

2. One piece of real material

Bring something that did not originate inside the model:

• an experience;

• an observation;

• an example;

• a result;

• a customer question;

• a source;

• a conversation;

• a lesson from your work.

You do not need all of them. One specific piece of material is more useful than a page of generic background.

3. Your rough interpretation

What do you think the material means?

The answer can be incomplete. It can even be one of several possible interpretations you are considering.

The important thing is that AI does not silently select the final interpretation while producing polished prose.

4. A reason the reader should care

What should change after someone reads the piece?

Should they understand a problem differently? Avoid a mistake? Make a decision? Try a process? Notice something they previously overlooked?

Without this decision, the draft may contain information without creating value.

Together, the framework looks like this:

The readiness test

Before drafting, try to complete this sentence:

Your answer does not need to sound good. It only needs to be honest enough to develop.

If you cannot complete it, the answer is not to force a draft. The answer is to keep thinking—and this is where AI becomes genuinely useful.

AI can help you find the point

The writer does not have to do all of this alone before opening an AI conversation.

AI can ask questions, identify tensions, challenge assumptions, suggest interpretations, and help connect material that initially feels unrelated.

The boundary is not between human thinking and AI involvement.

The boundary is between exploration and approval.

AI can help surface possible points. The writer must consciously accept, reject, or rewrite the point that becomes the basis of the draft.

That gives us a better workflow:

Instead of asking AI to write immediately, you make it interview you first.

The interview pulls the missing decisions into the open. The Content Map then shows you what the model believes the piece is about before that interpretation disappears inside polished prose.

You can correct the reader, claim, evidence, limitation, or structure. Only when the map reflects what you genuinely want to say do you approve it and allow drafting to begin.

Why this produces a stronger process

Google’s people-first content guidance asks whether content serves an intended audience, demonstrates firsthand expertise or depth, contributes original information or analysis, and helps the reader achieve a goal. Those are not merely SEO questions. They are useful tests for whether a piece contains something worth publishing.

This workflow applies the same principle to AI-assisted writing.

It does not prevent AI from helping with discovery. It prevents discovery, editorial approval, and drafting from collapsing into one invisible step.

Use AI for prose after the point becomes visible

Once the reader, material, meaning, and relevance are clear, AI can do a great deal of useful work.

It can organize the argument, test different openings, improve clarity, adapt the structure to the chosen format, and turn rough language into a usable draft.

But the sequence matters:

Below is the exact prompt that runs that entire process inside one conversation. You paste it once, provide whatever real material you have, answer the interview, approve the resulting Content Map, and receive the draft automatically.

The paid section includes the complete prompt, exactly what will happen when you use it, how to request different versions, and the final checks to run before publishing.

Here’s the exact prompt to turn your rough material into a strong piece of content with AI:

Read more

Every model learned to write from the same internet, so every model reaches for the same words. This is actually getting worse in some aspects since these same models are then being trained on this writing again and again. It’s a pretty vicious cycle.

“It’s not just X, it’s Y.” Readers have seen those moves ten thousand times, and the moment one shows up, many of them close the page. You know what’s worse? That page might even contain incredible insights, but the writer could never get around the AI tells.

One way you can start fighting this is by keeping a file that bans all of it. I pulled hundreds of words, phrases, and sentence patterns that readers actually complain about. I looked across several large studies and real editor complaints, including my own experience throughout the past couple years.

Every piece I publish now includes this as one of my many gates. I’m giving you that file.

One file bans the whole AI vocabulary

The file is plain text, organized so your AI can actually follow it and it includes:

The infamous words. Delve, tapestry, robust, seamless, pivotal, “dive into.” The ones with studies and headlines behind them.

The stock phrases. “In today’s fast-paced world.” “It’s important to note.” “That being said.” The openers, hedges, and transitions no human types.

The sentence patterns. The structures readers now spot instantly, starting with the single most recognized AI tell there is: “It’s not just X, it’s Y.”

The dead giveaways. “As an AI language model.” Chatbot leakage that ends careers. You search for these before anything ships, every time.

The false positives. The section most lists don’t have: the human things (your pet words, your rough edges, real lists, clean formal prose) that you should never let anyone edit out of you.

All you gotta do is point your AI at it.

Save this file in whatever you write with. Add it to your project files in Claude or ChatGPT, drop it in the folder you work out of with Claude Code or Codex, paste it into your instructions, or include it in any skill you build. Then tell your model one thing: check every draft against this file before showing me anything.

Your next draft comes back without the vocabulary everybody is sick of.

Download the AI-tells file here:

Read more

I want to push back on that this week, and I get to do it from an unusual position.

I know slop better than the people yelling about it. I used to make it by hand.

The job was volume, every day, by hand

Before the models existed and right after I graduated college, I was freelance writing to make money. Not work I really cared about, not a portfolio I was building. It was how I paid for things. I wrote for a couple of shady operations. One such operation was a blog about horse racing: races, results, betting angles, tracks across the United States and the rest of the world.

The routine never really changed. I would look at the race, read the various results and reports, write it up manually, move to the next one. It got boring fast, and the monotony was the job.

It became a brainless job of putting very little thought into any of it, basically copying and regurgitating the result of a race that had already happened. Then came the part the editors actually cared about: slap a sh*t ton of keywords into it and follow whatever SEO guidelines they sent over that week. The articles were never really meant for the readers. They were for Google. That’s what many of those early blogs were.

I got paid twelve to twenty dollars a piece. So the game was volume. Produce as many as you can, every day, by hand. (You would be surprised at how many horse races actually run daily.)

I was producing slop. I didn't really know it at the time. Or maybe more accurately, I just wasn't even paying attention. I was doing what I was told to do, and I was damn happy I had someone paying me to write.

I mean I barely knew the sport. My entire history with horse racing was one Kentucky Derby when I was sixteen: I made a few picks, my older brother placed the bet, and I hit the trifecta. That was it. That was the guy covering races all over the world. Was I an expert? Did I provide unique value? (Perhaps. I don't know any other 16-year-olds hitting trifectas.)

There was no AI anywhere in the building at the time. It's almost hard to remember what it was like before everything was AI.

The important thing is: slop existed long before the models did. People just didn’t call it out the way they do now. They didn't even know they were reading it. It went by different names: keyword stuffing, SEO manipulation, content farming. The internet was drowning in it a decade before ChatGPT, and nobody was scanning anything. People want to pretend slop was born the day the chatbots launched. I was making it years before the models could.

I asked my readers how they define slop

So the word needs defining, and this week I didn’t want to define it alone. On Wednesday I asked the WriterOps chat: is “AI slop” a real thing, or has the label stopped meaning anything?

There were two answers that I liked, and I think they pretty much sum up the two ways the world looks at AI slop.

The first definition is about value. 's version: slop is when the user “doesn’t input enough of their own organic material and lets the robot speak for them.” Nothing of the person in the work, so the machine fills the space with what it would say to anyone.

That definition holds up, and underneath it isn’t actually about AI at all. It's a user problem. A "skill problem,” if you will.

Slop is content with zero value. Volume with nothing baked into it and no unique experience, effort, editing, or finessing the model until it says what you actually meant. It's vanilla.

Isn't that the slop I was making in the horse-racing years, by hand? Even the dictionaries agree with her: Merriam-Webster made “slop” its word of the year for 2025, defined as low-quality digital content, produced in quantity. There’s no AI in that definition. And the fix follows straight from it: provide value, whatever it is you do.

You'll still get tagged, and that's fine

's response made me laugh. It does feel dirty. And for me, even worse is that it's just straight up annoying, and we have enough of that already online.

Now I said there is a mania over AI slop right now, and this is the one the mania runs on. This is slop as a texture. The tendencies: the words every model overuses, the phrasing patterns, every piece shaped like every other piece. In practice I’d argue this is the bigger definition. Most slop calls happen on sight, from the tendencies alone, way before anyone even asks whether the piece said anything. It signals to a lot of people to just disregard what's in front of them and keep scrolling.

And there is real work you can do about this dirty texture. Try feeding the machine your context instead of publishing its defaults. That's the least you can do.

Better yet, interview yourself and harvest your actual views into the system. Dictate your ideas out loud and have the AI organize them instead of inventing them. Keep reference documents that ban the language everybody’s model reaches for; mine block hundreds of words and phrases. This is a real craft. Don't let someone tell you otherwise.

Now the honest part. None of that will get you past the scanners. If the AI is doing the writing, the detectors will tag it, and no amount of technique changes that.

And that is perfectly okay. Roughly half of new articles on the web are now primarily AI-generated. You think every one is worthless? Of course not. If a tag defined slop, half the internet would already be worthless by definition, the argument would be over, and everyone would have lost. A detector can tell you a model was involved. It can’t tell you whether the piece was worth your attention, and it never will, because that was never a machine-readable property.

Don’t let people put you down over a scan result. And they will try. If you’re not a fiction writer, if your essays aren’t meant to be philosophy, if you’re not a Dan Koe and were never trying to be, then some AI tendencies are an acceptable trade for finally communicating well. Writing, for most of the people doing it, is a form of communication. Do what you can about the language. And make sure, above everything, that the work carries an idea or a perspective that isn’t replicable by everybody else. That standard survives both definitions.

The operation decides what the writing is worth

Back in the SEO era there were two directions you could take, and I lived both of them.

The first was those shady agencies. They didn’t give a f*ck about the writers, paid us almost nothing, buried us in keyword guidelines, and made us re-edit finished pieces every time Google changed its mind about what ranked. The work didn’t even carry our names. The articles went out under the CEO’s byline, under the leadership’s names. You could call it ghostwriting if you want. I don't. To me it was something else and those businesses didn't last. I watched as every day was a struggle. It didn't look too fun to me, and the arrival of AI only finished what their own model had started. I quit before the end. And I was stressed and broke.

The second direction was the respectable publications I found my way into afterward. Their model was the opposite in every way that mattered. Make the content unique, make it valuable, give the reader something, write from authority and a point of view, and let that do the ranking. Reader over everything. The rest will follow.

AND they put my name on the work. Encouraged it, actually. Wanted the writer visible and proud of the piece, which made me care more about it in return. Those operations blew up into millions upon millions of views a month, and working inside them launched my career. Same writer in both directions. The difference was never the writer. It was what the operation cared about.

Those same two directions exist today, with much better tools.

One group uses AI the way my old bosses used keywords: pump the feeds, game the algorithm, volume with nothing inside it. That’s the equivalent of keyword stuffing today, and I expect it to end the same way, because it already ended that way once.

The other group runs AI as an operating system. The model might be the communicator between me and you, but everything it says passes through a system first. The final writing, even when you can pick up AI tendencies in it, comes from somewhere a generic chat session can’t reach. That’s the actual difference between the two directions. It lies in what the machine was connected to, not whether there was a machine to begin with.

And I want to be precise about where the line sits, because I run high-volume operations myself. Volume was never the crime. Scaled content isn’t slop by definition, any more than one precious hand-typed essay a month is value by definition. Some of the most useful work AI does right now only exists at scale: processing more sources, more data, and more developments than any human could hold in their head, and surfacing connections a person staring at one corner of the field would never see. Getting models to do that reliably, and keeping them on track while they do it, is its own discipline, and almost nobody has really tapped it yet.

The point of view doesn’t have to live in every sentence. It can live in the system: where you point the machines, what you force them to verify, which connections you teach them to hunt for, what you refuse to publish.

The dreaded scan

The industry is reacting to this flood, and it’s mostly aiming at the tool. They always blame the tool, never the user. In the last month alone: Substack gave readers a Pangram scan button for every post over a hundred words, the EU began requiring disclosure of AI-generated public-interest text unless a human holds editorial responsibility for it, the New York Post launched an AI named Hamilton that is explicitly forbidden from writing the journalism, and Anthropic started watermarking the words Claude generates. Four institutions in four weeks, all building machinery around one question: was an AI involved?

When the scan showed up on this platform, my first reaction was that it’s a gimmick. A few weeks in, I’ve moved some. I don’t hate it, and I’m not against it. I actually think Substack is doing better than most when it comes to fighting AI slop while providing users some control over the process.

The bigger problem is the mania and gatekeeping you see among certain groups. “AI-assisted” should open a question: did this person put anything of themselves into the work? Even though Substack does a good job of letting writers explain their process, I don't know if the people who are most obsessed with this actually care. They see AI, they keep scrolling.

I think it's important to realize that the reflex didn’t come from nowhere. It was the combination of actual AI slop that nobody can handle anymore, and a lot of loud voices in media spending the last two years demonizing anyone who used AI.

I digress. We will cover that another day. Everything I do is based around implementing AI into writing operations (WriterOps). So you know where I stand.

But there are two bigger arguments hiding under this one, and each deserves its own essay later.

The first is the double standard. We hand AI real credit everywhere else: biology, medicine, nutrition, finance, investing, legal work, real estate, images. Nobody dismisses a drug candidate because a model did the folding. Writing is the one domain where the public gives AI zero benefit of the doubt, and nobody stops to ask why the rules changed at the paragraph.

The second is the flip side of the whole panic, and it’s the one I care most about. For the first time, a person with a genius point of view and no writing ability can actually communicate it to the world. That knowledge used to stay locked inside people who couldn’t write it down. Now it doesn’t have to. That is immensely valuable, and right now it gets zero credit. This is a democratization of communication. And if you look closely, you start to see the same pattern we see all over the place when it comes to gatekeeping an industry or a skill.

Do these three things today

If you use AI in your writing operations and this mania has you flinching, here’s where to start.

1. Run the mirror test. Ask what your AI setup actually knows about you. If the answer is nothing, then its output would look identical coming from anyone else on earth. I'm sorry to tell you, but that is the definition of slop. And you're not gonna get any empathy from me or anyone else. Start feeding the machine pieces of yourself or your brand right now.

2. Ask the pre-publish question. Before anything ships: what’s in this that only I could have written? If you can point to it, the value proves itself. If you can’t, don’t publish it. And if you run an operation instead of a byline, ask it at the system level: what’s in here that only this operation could have produced?

3. Fix your posture. Whatever you do, don't apologize for the tool. The shame belongs to valueless output, and it always has, since long before the machines showed up. Use AI. Hold your standard. Put your name on the result.

That last one took me the longest, and I learned it in reverse. I spent years producing work that nobody would sign, least of all me. I know exactly what that work was worth.

The scan can read the words. It can’t read whether you did the work.

That part’s yours.

—Alex