Almost nobody talks about the specific category of thing it can never do, at any level of sophistication, no matter how much compute you throw at it. Not because the systems are not smart enough. Because the category itself is structurally inaccessible from a training set. And that category is the entire foundation of every piece of work you produce that anyone actually wants to read.

Your interior life.

The specific experiences you have had. The specific way those experiences arranged themselves into the way you now see the world. The tastes you developed by living inside a particular body for a particular number of years. The scars, the intuitions, the specific attention to specific things that came from the specific life you actually lived. None of this exists inside any model. None of it is available for the model to reproduce. And it is, precisely, the thing your best work has always come from.

You have been worried about AI outsourcing your work. What you should notice is that the parts of your work that are outsourceable were never your work in the first place.

Subscribe now

What AI can and cannot access

An AI can process every text ever written. It can identify patterns in language, structure, argument, form. It can produce reasonable summaries of positions humans have already articulated. It can recombine existing material at speed.

It cannot access what has not been written yet.

Every specific insight you have that you have not yet put into words is unavailable to any AI in existence. The system can approximate what similar people have said about similar topics, based on their public writing. It cannot approximate what you would say, because what you would say is a function of who you specifically are, which is a function of a life the model has no visibility into.

This is the specific human moat. Not intelligence in general. Not craft in general. The specific interior of a specific person, arriving on the page as a specific claim shaped by a specific history. That combination is not a training-set problem. It is a source problem. And the source is you.

The mistake most creators are making

They are using AI to outsource the parts of their work that came from their interior, while keeping the parts that could have been done by anyone.

The interior contributions get delegated: what they think, what they notice, what they claim, what stance they take. The mechanical parts stay: research summaries, formatting, minor edits. This is exactly backwards. AI is excellent at the mechanical parts. AI cannot access the interior parts at all. If you are letting the tool handle your positions and taking on the formatting yourself, you have inverted the entire relationship.

The correct relationship is the reverse. Keep the interior contributions ferociously to yourself. What you notice, what you think, what you argue for, what you are willing to defend. Outsource the mechanical work around those contributions to the tool.

The people producing excellent AI-assisted work are doing exactly this. They are letting the tool handle what the tool is good at. They are protecting, aggressively, the specific human input the tool cannot generate. And their output feels sharper for having a machine in the loop, because the machine took over the parts that were slowing them down while they kept the parts that make the work worth reading.

The move

For your next piece of work, run one specific audit.

Identify the interior contribution. What is your specific claim, your specific noticing, your specific position, your specific stance? That is your contribution. Nothing about it should touch an AI system. It is the thing only you can produce.

Then identify the mechanical work around the contribution. Structure, phrasing, minor edits, format. Let the tool handle those, where useful. Let it be fast at the parts where speed does not compromise anything.

You will notice something quickly. The pieces where you did this cleanly have a specific quality your fully-AI-produced pieces do not have. That quality is you, present in the work, uncorrupted by outsourcing.

The tool is a lever. It multiplies whatever you put into it. If you put in your interior, you multiply your interior. If you put in nothing, you multiply nothing.

The interior is the only thing you have that nobody else has. Do not let a machine be the one that keeps it for you.

You open an AI tool. You type in a prompt. The output comes back generic, uninspired, exactly the kind of thing that would embarrass you if you passed it off as your own. You blame the tool. Maybe the model. Maybe the training data. You try a different tool. Same result. You conclude that AI is overrated.

There is another explanation you have not seriously considered. The problem might not be the tool. The problem might be the prompt. And the problem with the prompt might be the thinking behind the prompt.

AI is not a magic system that produces good work from any input. It is closer to a very fast collaborator who can only respond to what you actually give it. If your input is vague, the output will be vague. If your input reveals that you have not yet worked out what you actually want, the output will confirm that you have not yet worked out what you actually want. The output is not failing you. It is showing you exactly what you asked for.

Which means: the specific quality of your AI output is a reasonably honest report card on the specific quality of your thinking.

Subscribe now

Why this feels insulting

It feels insulting because it moves the failure from the tool to you.

If AI is the problem, you are off the hook. The tool is broken. You can complain about it, ignore it, dismiss it. Your own thinking gets to stay uninvestigated.

If your thinking is the problem, that is a different kind of situation. It means every mediocre output you have produced was accurate reporting on the specificity of your input. Which means the vagueness of your outputs has been telling you something about the vagueness of your ideas the entire time, and you have been blaming the messenger.

The people who consistently produce excellent AI outputs are not people with better prompts. They are people who have done the harder underlying work of getting clear on what they actually want, what the actual problem is, what would count as a good answer to the actual question. The prompt is downstream of that clarity. If the clarity does not exist, no prompting technique compensates.

The specific tell

Here is a diagnostic worth running.

Take one of your recent AI outputs that disappointed you. Read the original prompt again. Ask honestly whether the prompt made it clear what you actually wanted.

Most of the time, when the output disappointed you, the prompt was vague. It described a topic rather than a request. It gave direction without specifying constraints. It asked for a thing without naming what would count as a good version of that thing. The prompt was, in effect, a poorly-formed thought handed to a system that had no choice but to produce a poorly-formed response.

You are not disappointed with AI. You are disappointed with your own thinking made visible.

For most people, this is uncomfortable in a way they did not sign up for. They thought they were adopting a productivity tool. What they got was a mirror that shows them exactly how developed or underdeveloped their thinking actually is. And most people, when they look in that mirror, discover they were operating at a lower resolution than they assumed.

The move

Before you write another prompt, write out the thing you actually want, in full sentences, before compressing it into a prompt.

Not the topic. The specific thing. What is the request? What are the constraints? What would count as a good version of the answer? What would you reject as unhelpful? If you cannot answer these questions in prose, you are not ready to prompt. You are ready to think.

Then, once the thinking is clear, the prompt writes itself. The prompt becomes short because the thinking behind it is long. The output improves not because you learned a technique, but because the input finally reflected work.

AI did not become useful for the people who mastered prompting. It became useful for the people who used the tool to notice, painfully, how vague their own thinking had been.

The mirror is uncomfortable. The mirror is also the whole point.

You didn’t need to be the best in your field. You just needed to be reliably competent - better than the bottom, close enough to the top. The middle of the bell curve was a comfortable place to live. It rewarded consistency, not brilliance. You got paid, promoted, and respected for producing solid work at a professional level.

That entire strategy has been eliminated in the last two years.

Not because AI is better than the top of your field. It isn’t. It’s because AI is now better than the middle of your field. The competent, professional, generic-but-well-executed output that used to define most of the economy is now available on demand, at zero marginal cost, produced by a machine that doesn’t take vacations or need onboarding.

The middle of the bell curve - where most jobs, most content, most work has lived for decades - just disappeared as a viable position. Not overnight. But close.

Subscribe now

What used to define the middle

Middle-of-the-curve work was work that required competence but not distinction.

A well-executed marketing email. A professional article on a familiar topic. A functional line of code doing something a thousand people had already coded before. A report summarizing what everyone already knew, packaged well enough to circulate. A design that met specifications without saying anything specific.

This was the volume of the economy. Most people did this kind of work. Most people were paid for this kind of work. And the strategy was to be good at it - good enough that people would keep paying you.

The strategy worked because generic-competent work required a human. The knowledge was replicable; the execution wasn’t automated. So competent humans commanded competent wages, and the middle was livable.

Now the execution is automated. Which means competent generic work has become approximately free - and the humans who produced it professionally have a specific structural problem they haven’t fully absorbed yet.

What survived at the top

The top of the bell curve did not shrink. It expanded.

The top is where taste lives. Judgment. Specific experience. The particular way of seeing that only comes from having lived a specific life and worked at a specific craft for years. The willingness to hold a position that could be wrong. The interior work of having something to say that isn’t a recombination of what everyone else is already saying.

None of this is automated. None of it will be automated in the near future. Because it isn’t a matter of processing existing information more efficiently. It’s a matter of producing new specificity - and specificity, by definition, cannot be generated from a training set of everyone else’s outputs.

The people at the top are, if anything, more valuable now than they were two years ago. Their scarcity became visible when the middle collapsed. Which means the winners of the AI age are not the people who learned the tools fastest. They’re the people who had already been developing what the tools cannot do.

The bottom still exists too

Physical work, embodied work, real-world work in messy environments - this hasn’t been touched. A plumber is still a plumber. A nurse is still a nurse. The floor of the economy remained where it was.

What got eliminated wasn’t the top or the bottom. It was the specific middle position of cognitive-but-generic - the desk-based, competent, professional work that used to sustain most of the middle class.

If that’s where you’ve been living, you have a real problem. Not tomorrow. Now. And the problem is not that you’re going to be replaced. It’s that you’re going to be commodified - your work will become interchangeable with a machine’s, and you’ll be priced accordingly.

The move

Look at the specific work you do, honestly, and ask: is any of this so generic that a machine could produce it acceptably?

If yes, that work is on its way out of your job description. Not because you’ll be fired. Because the market will stop paying for it.

Move upward. Not by working harder at generic work. By developing the specific taste, judgment, and interior that constitutes top-of-curve work - the work AI cannot do, because it requires having lived a life the machine hasn’t lived.

The middle used to be a livable position. It isn’t anymore.

The top is where you now have to live to be paid like a human.

What part of your current work would you replace with AI if you didn’t have to admit it - and what does that tell you about where you need to move?

– Hasnain