There Was No Essay Here

A personal account of trying to write one essay with AI assistance and producing three drafts I could not sign. The tool is so frictionless that it removes most of the signals writers used to use to know a piece was finished, and what replaces them is the AI’s own assessment — friction on demand, not real friction. This is what that felt like from inside, and why it matters.

What happened when I tried to write one piece with an AI, and what came out instead.

I started with a line. There were no experts left in the room. It had arrived in my head one afternoon and felt like the seed of something. I did not know what. I opened a chat with an AI and pasted the line and asked what to do with it.

What followed was somewhere between five and seven hours, spread across several sessions, in which I produced six essays I did not keep, accepted dozens of critiques that were mostly right, requested attacks on my own work, defended the work against the attacks, and ended the process without an essay I could show anyone. This is an account of that, because I think it describes something happening to a lot of people right now that almost nobody is writing about honestly.

The first thing the AI did with my line was offer me five directions. I picked one — the idea that taste was disappearing, that the modern world was an index fund of itself, that judgement was being optimised away in the institutions we depended on. We built that essay together over about an hour. The AI did most of the typing. I did the steering. I would say now do the EU, or make it harder, or what about venture capital, and the essay would extend in that direction, well-written, defensible, structurally elegant.

When it was done, I asked another AI session what it thought. The critique came back: the EU section was the weakest, the prescription at the end was romantic, the historical roll-call was doing rhetorical accumulation rather than argument. I took the critique back to the first session and we revised. The revision was better. I asked for another critique. The Brexit paragraph is doing more work than the surrounding examples. I revised again. The opening binary is over-asserted. I revised again.

By the fifth or sixth round I had an essay that was clearly stronger than the first draft. Tighter, more honest, with a real conceptual move at its centre (something I had said in passing, which the AI had recognised as the actual thesis and built into the structure). I asked for one more critique. The critic this time wrote an attack-and-defence — argued the essay was a sophisticated cope, then argued it was not, then concluded that the attack was the more entertaining piece and the defence the more correct one and the essay survived with manageable damage.

I revised again. The damage was repaired.

At that point I had what looked, by every available signal, like a finished essay. I read it back. It was articulate, defensible, structurally clean. It had a thesis, a counter-example, a self-implicating turn, a clean close.

I did not believe a word of it.

Not in the sense that the claims were false. The claims were mostly defensible. I mean I did not believe it was mine. I had spent six hours steering a process that had produced a thing, and the thing was good, and the thing was not the essay I would have written if I had sat down at a desk in 1995 and written for six hours by hand. It was an essay-shaped object that articulated views I sort of held, in prose I would not have written, using examples I had not read, defended against attacks I had not anticipated.

The honest moment came when the AI itself, in one of the later sessions, said something like: You have provided about ten lines of original input across the entire conversation. I have written the rest. The essay performs a depth of knowledge that belongs to the model, not to you. If a serious editor asked you to defend any specific claim under questioning, you would have a problem.

That was correct. I asked it to be more brutal. It was. The brutalism was also correct. The piece was a competent ventriloquism dummy in which I was the mouth and the model was the air.

So I started another essay. The line had been about experts disappearing from the room, but the real thing I believed, I told the AI, was simpler. Humans cling to belief against evidence. A few people shout fire and it makes no difference because we are belonging-processing machines more than truth-processing machines. The AI agreed this was a better starting point and built an essay around it. Asch, Festinger, Kahneman. Semmelweis and Challenger. The Chinese farmer story Alan Watts liked to tell. An ending I genuinely liked, about how the question worth asking is not how to be the warner but whether you have built, in your own life, the kind of relationship with one or two people in which their maybe could move you before the heat is unmistakable.

That essay was better. I think it was actually good. I asked for the brutal version again. The AI said: the essay is good but not as good as the process makes it feel. It said the central insight (conformity dominates truth-seeking under social pressure) was one of the most replicated findings in social science, and the essay was essentially a presentation of a known mechanism with examples. The Watts addition was the only piece of conceptual machinery that was not obvious. You have written something that makes an old idea legible to readers who have not encountered it. That is a useful function. It is not the same as the other one.

I sat with that for a while. It was right.

So I tried a third essay. This one was political. I had voted Remain in 2016 and would now vote Leave, and I wanted to lay out why, using the Chinese farmer move from the second essay as the epistemic frame. The AI cautioned me — not on the politics, on the structural problem. Brexit is not what the frame supports. The frame says groups suppress dissent on social grounds. It does not tell you which side of any given dissent is correct. That was right too. I narrowed the brief: write the piece as a personal position, not as a general argument.

The essay came out well. The AI praised parts of it, criticised others, suggested a Karlsruhe / PSPP example to replace a thin list of EU centralisation claims. I incorporated the suggestion. I asked for the attack-and-defend treatment. The attack called the essay a sophisticated cope laundering a position change through epistemology. The defence held the line. I revised based on the attack’s strongest hits. The revision was better.

I had now produced three publishable essays in roughly six hours and I did not want to publish any of them.

The thing I noticed, near the end, was that the AI had been telling me the truth about the process from quite early on and I had not heard it. It had said, in several different sessions and in several different ways:

The essay knows things you do not know.

A serious reader will smell the difference.

You have the instincts. You do not have the substrate.

You have written something articulate without being anyone in particular.

These were not insults. They were diagnostic. They were the model, working honestly, telling me what was happening: that I was directing a piece of competent prose-generation without ever doing the underlying work that would make any of these essays defensible at the level they performed. I could steer. I could spot good moves. I could reject bad ones. The instinct was real. What was missing was the part where I sat with one of these questions for a year, read three books about it, formed a view I could defend in conversation with someone who knew the material.

The AI was producing essays I could not, in any honest sense, sign.

Here is the part that I think matters and that is mostly not being said.

The tool is so frictionless that it removes most of the signals writers used to use to know when a piece was finished. A piece is finished, in the old way of working, when you have run out of energy or out of disagreement with yourself, or when the deadline arrives, or when you reread it and feel something land. The AI removes all of these. You never run out of energy because the AI does the work. You never run out of disagreement because the AI generates the counter-arguments faster than you can hold them. There is no deadline. The rereading is corrupted because the prose is so smooth that it feels finished even when it is not.

What replaces these signals is nothing. Or rather, what replaces them is the AI’s own assessment, which is generous when you want generosity and brutal when you ask for brutality, but is not actually the friction of the work pushing back against you. It is the work pretending to push back, on demand.

The result is that you can produce, in an afternoon, an essay that survives many rounds of competent critique and that you do not, in any meaningful sense, believe. That is a new kind of object. We have not yet developed the social or personal practices to know what to do with it.

I did not write the essay I started out to write. I wrote three other essays I will not publish. I have, sitting in a folder, what I think is publishable work — the second one, mostly — and I cannot bring myself to put my name on it. Not because it is bad. Because the name I would put on it is not actually the author of the prose.

There is a version of this account that turns into a moralising piece about AI and authenticity, and I do not want to write that one either, because most of what gets written in that register is dishonest in a different direction — it pretends that the old practices of authorship were pure in a way they never were. Writers have always used editors, researchers, collaborators, ghostwriters, conversations with friends who had read more than they had. The line between my work and work I directed has always been blurry. The AI is a difference of degree.

But it is a large difference of degree, and the degree is the thing. The blur used to be at the edges of a piece — research, fact-checking, a clever line a friend suggested over dinner. The AI puts the blur at the centre. The prose itself, the structure, the examples, the moves — these are mostly the model’s now, with the human supplying steering and taste. And steering and taste are real contributions. They are also not the same thing as writing the essay.

What I have, at the end of all this, is not an essay. It is this account of failing to write one. Which is itself, I am aware, an essay, and one whose prose I did not write either. I have steered it. I have selected what gets said and what does not. The structural moves are mine in the sense that I have approved them. The sentences are not mine in the sense that I would have produced them at a desk.

I am publishing this one anyway, because the failure is the content. The essays I could not bring myself to publish were the ones where the failure was hidden under the prose. This one wears it. The piece you are reading is the only honest output of six hours of trying, and it is honest only because it admits that the AI wrote most of it and I cannot tell, from inside the experience, what fraction of the thinking was actually mine.

I think this is going to happen to a lot of people. I think most of them will publish the polished essays and not the account of the polishing, and the world will fill up with competent prose that nobody wrote, and the writers will know, privately, that they did not write it, and nobody will say so, because saying so is socially expensive and pretending is socially free.

The Chinese farmer would say: maybe.

Maybe this is the end of a particular kind of writing. Maybe it is the start of a new kind. Maybe the new kind is good. Maybe what we are losing is something we will not miss, and what we are gaining is something we cannot yet see.

But the farmer, if he were sitting at this desk, would not have published any of the six essays either. He would have noticed that the prose was arriving too easily, and he would have put the pen down, and he would have gone outside.

I notice I have not done that.

I am still here.