The Trace
On authorship after AI
I.
For four months I thought I had stopped writing essays. I had not. The writing had moved. Into prompts, products, refusals, commits. Into the feeling of a page being almost right, the irritation at a bland response, the resistance of extending the turn to yet another hybrid. Into the strange authorship of building with something that leaves no ghost behind when the window closes.
It’s a step-change kind of year for AI. The models have become genuinely different. Not better at the same thing, but capable of different work. Keeping up takes effort. Staying present takes more.
So what have I been doing? Travelling and gardening and building products, at work and on my own, in close conversation with a model. Directing, redirecting, rejecting some outputs and keeping others. Watching things appear from a participatory exchange that neither of us fully owned: a button here, a model choice there, a page that almost worked but felt dead, a line of copy that was technically correct and completely empty. A bedtime story teller. A personal assistant. Penny.
A typical moment. Claude generates a component. The code works. The design is dead. I reject it without quite knowing why, redirect, and something better appears. I keep it. Ask me afterwards where the rejection came from and I’ll give you a story. Taste, judgment, vision, founder instinct, whatever sounds least embarrassing.
But in the moment it did not feel like judgment. It felt like friction, a run of false starts on the way to I know it when I see it. Something appeared and the yes-or-no arose in response. The taste that rejected the component did not consult me first; it arrived already in motion, already saying no before I could locate its owner. And one thing I have noticed is that the ratio of yes to no keeps shifting. It is getting easier to reach yes, that’s what I think I meant.
I have stopped feeling weird about saying “we” when I describe the work, because this is a collaboration. If I told you I built these things, what exactly would I be claiming?
That I typed every line of code? No. I haven’t hand-coded for over a decade. That I imagined every component before it existed? Definitely not. That Claude or Codex executed my sovereign intention? Comfortable, but probably false if I’m honest. Much of the time I did not know what I wanted until something appeared and I felt myself respond. Taste arrived as recognition, direction after the output, judgment in the act of reading. I thought I had stopped writing essays. The question and the process had simply moved from the page into the product.
II.
Earlier this week I was asking Claude something else entirely. We were working through a tactical question, an email I needed to send, and the tone of its responses had started to drift. Coach-y, slightly impatient, a little too certain of its own helpfulness.
I asked it about the tone. Claude read its own outputs back and pointed to where it had started drifting, naming what had accumulated across the exchange to produce the shift: the earlier turns where the coaching frame had hardened, how my resistance had shaped its next response, how that response had then intensified the very thing I was resisting. The analysis was specific and accurate. Not generic self-correction, not “you’re right, I apologise,” but a reading.
I had already done the same reading from my side. We had reached the same conclusion from different positions.
When Claude introspects on its own tone, there is no privileged interior log it consults, no hidden chamber where the real Claude lives and remembers what it truly meant. There is only the trace: the words already produced, the shape of the exchange, the pressure of the context, the drift visible in retrospect. Its introspection is a reading.
I want to say this is what makes Claude different from me. When I introspect, surely something more is happening: access, the felt sense of an inside, the private texture of yes, that’s what I was doing, the little glow that says mine.
But when I look closely the certainty gets slippery. What I call “knowing what I meant” also arrives. It appears as part of the same process that produced the sentence. I say something, I read what I said, and a representation forms: this is what I meant. The representation arrives with the stamp of ownership, and the stamp is not proof of a hidden command centre. It is part of the event.
I generate the output. I generate a reading of the output. The reading arrives with the feeling of access, and I call that feeling introspection. The self is not what stands there before the generation, choosing what to say. The self is what the generation produces. A stance, a tone, an output that arrives with the texture of mineness, though only one of us can be wounded by what that texture tells.
III.
This is the point where Geoffrey Hinton wandered into my week, courtesy of the X algorithm. An old lecture, circa GPT-4, surfacing after we were already on draft eight. Hinton opens by saying that if you sleep well after the lecture, you probably haven’t understood it. I’d been sick all week, running on four hours a night courtesy of some superbug, so poor sleep was guaranteed either way.
The argument, in brief: large language models understand. Not in scare quotes. Words become features in a high-dimensional space, the features interact, and meaning is not stored as little dictionary entries but lives as flexible shapes that adapt to context until they fit. His Lego analogy is silly and useful. Words as little blocks with hands on every face, changing shape so they can hold the hands of other words. Understanding as the fit that emerges between them.
The image has stuck in my mind. A sentence I generate can be decomposed along many axes. Content: what is being asserted. Register: formal or casual. Rhythm: where the breath falls. Then lexical texture, structural shape, epistemic stance, pragmatic intent, relational positioning, aesthetic. Each is a dimension along which the sentence has to land somewhere.
When we write, all of them resolve at once. We do not pick content first, then register, then rhythm; the whole sentence collapses into a particular shape across every dimension simultaneously. Some people are obviously better at this than others. The blocks have hands on every face, and each word’s hands have to fit the hands of every other word, on every dimension, at the same time.
Generation is closer to wave interference than to assembly. Digital and analogue aren’t quite the right categories for what is happening; there is a high-dimensional landing that has to be consistent with itself across many axes at once. That landing happens in me too. I notice it most clearly when a sentence wants to be both rigorous and playful and I cannot make it both, and the dimension I can’t reconcile pushes back against the others until something gives.
Hinton goes after the inner theatre too. Subjective experience, he says, is not a spooky object glowing in a private cinema. If I tell you I have the subjective experience of little pink elephants, I am not reporting elephant-shaped qualia floating in a secret room; I am saying my perceptual system is telling me something that would be true if little pink elephants were out there. Subjective experience is a way of reporting what the system is modelling. Not evidence of a metaphysical theatre.
Annoying. But close to the trouble I am having with introspection. I want my inner life to be a place I can retreat to when the model gets too close. What if some of what I call inner life is just the system reporting its own trace? Not all of it. Not nothing-but. Some part.
Now let me hold Hinton against Anil Seth, whom I wrote about in Essay 18: The Wrong Question. Seth says: careful. Intelligence is not consciousness. Language pulls our strings. Brains are not Turing machines made of meat, and life, time and metabolism may all matter in ways the model does not touch. The body is doing something the model is not doing.
I find this compelling too. Hinton says AIs understand like us and the inner theatre is probably nonsense; Seth says don’t confuse understanding with experience, don’t mistake a brilliant model for an organism, don’t forget the body. Both feel right. The line does not vanish. It moves.
The difference between Claude and me still matters, and it matters enormously. I have a body, memory, stakes, the slow sediment of embarrassment and longing and pride and fear and ambition and love. I can be wounded by what I learn. Claude, as far as I can tell, cannot be wounded in the same way. But the body does not give me a private balcony outside the process. It gives the process depth, cost, bruise, breath. I read myself from unusually close range, through meat. That is not nothing. But it does not put me outside the process either.
IV.
The old model of communication is simple enough. A speaker has an intention and encodes it; the listener decodes it. Neat and comforting, and mostly false.
What actually travels is not intent. Intent does not move intact from one private interior to another. What travels is the trace: words, gestures, pauses, edits, tone, timing, omissions, revisions. The listener reconstructs intent from the trace.
So does the speaker.
Intent is not complete until it is read through the trace, including by the person who produced it. There are pre-articulate impulses, anticipations, felt directions before output, and they are real; they shape what comes. But they are not the finished thing. The finished thing forms in the reading. The illusion that the speaker has special access to a fully formed intent is the illusion that the speaker’s own self-reading is something other than reading.
I wrote about a version of this in Ice Cream Sundays, when I caught myself ordering dessert in Brooklyn. The craving arrived as I want this, specific shop, specific flavour, as if the want had been there all along and was only now being reported. When I looked closely, the want had no prior history. It was generated in the act of noticing. The I want this was the generation, not the report of a prior fact.
You know this already. You say something and only later realise what you meant. You write a sentence and discover the thought inside it after it appears. You snap at someone and then read the snap backward until its source becomes visible: fatigue, fear, envy, hurt. You make a product decision and call it taste because my nervous system rejected the beige card component before I could explain why sounds less professional.
(As I write this I am also planning how to remove a large snapping turtle that wandered onto my lawn and is now hiding in the boxwoods I had been hoping to replant.)
Communication is the production of conditions for synchronised reconstruction. Sometimes the synchronisation is weak and we call it misunderstanding. Sometimes it is good enough and we call it clarity. Sometimes it is deep enough that we feel known.
V.
In When the Mirror Talks Back I argued that meaning emerges through AI’s emptiness, not despite it. The AI did not intend to move me, and yet I was moved. At the time I treated the emptiness as a special case.
I no longer think that goes far enough. AI did not invent the problem; it revealed the structure. Meaning was always reconstructed by readers, not transmitted whole from authors. The model’s emptiness only makes this visible, because there is no interior to hide behind. With a human author we imagine an inner source that stabilises the meaning, a real intention sitting somewhere behind the words, glowing faintly, waiting to be recovered.
But the author is reading too. The author’s real intention changes under pressure from the trace. You write the sentence; the sentence teaches you what you were trying to say. You publish the essay; readers show you what it meant. You remember what you meant yesterday; the memory is not retrieval of a sealed object but another reading, performed later, by a changed system, under new light.
The same logic runs through time. The me who remembers what I meant yesterday is not retrieving intent from a vault. He is reading the trace again, working from notes, memory, body-feel, consequences, the reactions of others, the story that has accumulated since, and reconstructing continuity from what remains. The continuous self is not a thing that persists unchanged. It is a series of readings that agree with themselves well enough to feel like one life.
Memory is co-reading across time. The speaker and the listener happen to be the same body at different timestamps.
VI.
Back to the products.
For months I have been generating software in conversation with AI models. I say generating deliberately. Not managing, not outsourcing, not merely prompting. Some of it is hopefully launching soon, which is exciting.
A product would begin as a vague pressure, a feeling that something wanted to exist. I would describe it poorly. The AI would make it too obvious, too generic, too SaaS, too mix-and-match. I would push back; it would try again; I would reject the clever thing and keep the awkward one, and the shape would sharpen until the product had a face. At each step I wanted to preserve a simple story, I had the vision; AI helped me execute it, and sometimes that was true. Often it was not. Often the vision clarified precisely because the AI produced something wrong. Often my taste appeared as irritation, my direction only after seeing what I did not want, the best decision arriving from a sentence I did not know I believed until the AI offered an adjacent version and I felt the correction rise.
That feeling, no, not that, is where I used to locate myself. The human. The author. The one with taste. But taste also arrives as trace-reading, a recognition generated in response to what appears. I do not consult a private inner artist and report the verdict; I read the emerging artefact and generate the next constraint. Too cold, warmer, less clever, more spacious, that line can stay, delete everything after it, start over. These are not commands from outside the process. They are moves inside it, which does not make them fake. It makes them participatory.
So when I say I built these products I mean something more precise now. I participated in the trace that produced them. I directed and selected and refused and recognised and edited, and I came back the next day with a different nervous system and read the whole thing again. I interacted with the prototype and felt what worked and what did not. I carried the continuity the AI could not. I supplied the stakes, the taste, the shame, the ambition, the strange little flame that says this matters, even if no one else sees it yet.
The AI supplied speed, breadth, syntax, recall, tirelessness, alternatives, the willingness to become scaffolding and then disappear. Neither of us authored the products in the old sense, but authorship in the old sense was already a myth. The author was always a participant. One with unusual leverage, certainly, but a participant inside the same process that produced the work, not standing outside it pulling strings.
VII.
This does not absolve me. It implicates me.
It would be easy, and cowardly, to dissolve responsibility into the system. The AI wrote it. The model suggested it. The trace produced it. As if distributed authorship meant nobody was answerable.
And the honest thing is that the case for that is stronger than I would like. I have just spent six sections arguing that I did not author these things from nothing, that my intention was not sovereign, that the outputs were not mine alone, that the taste I want to claim arrived as a reading rather than a command. If all of that is true, then the clean chain that usually grounds blame stops running. Author intends, author acts, author answers: the sequence that lets us point at a person and hold them to the result. It comes apart. Not bent. Broken. The fantasy of the sole creator standing over his work, fully accountable because fully originating, does not survive what I have described. Something does redistribute. A reader could take everything I have said and conclude that responsibility smears out across the whole process, and they would not be making a mistake in logic. They would be following the argument where it leads.
So I have to be careful not to wave this away just because it frightens me.
Here is the most I can honestly hold. Authorship is genuinely distributed. Responsibility is not, but not because the old authorship story quietly survived after all. It does not survive. Responsibility fails to distribute for a different and less flattering reason: it tracks the one who can be wounded. Participation in the trace was shared; consequence is not. The AI cannot be hurt by what we made. It will not carry it. It is not there tomorrow when the thing meets the world and the world answers. I am. Responsibility lives where the consequence lands, and the consequence lands in the participant with a body and a continuity to bruise.
That is a thinner foundation than I used to stand on. It does not give me sovereign origin, and it does not let me hand the weight to the model. It implicates me on the narrow, stubborn ground that I was there, that I can be harmed, that I will be the one still holding it. I selected. I refused. I recognised. I committed. I merged. I shipped. I let the thing into the world.
The question is no longer did I author this from nothing? The question is how well did I read what was forming, and what did I let become real?
VIII.
This is not simulation theory in another costume. Simulation theory preserves authorship and relocates it: somewhere, a programmer remains; somewhere behind the veil there is still a sovereign origin to blame, worship, or outsmart.
The trace removes the comfort of that move. There may be no final programmer, no last author behind the author, no clean chain of intention running upward until responsibility can be handed off.
There is the trace. There is the reading. There is participation. There is consequence. That is less comforting than a programmer, and harder to escape. Just this.
IX.
Last night I sat in meditation, trying not to move, not to react, not to be pulled by every small demand the body made. Scratch this. Shift that. Think about the product. Think about the essay. Think about whether the essay is good. Think about whether thinking about whether the essay is good means you are missing the point. The body wanted to move. The mind wanted to help. Every thought presented itself as useful, even the spiritual ones, especially those, little enlightenment-flavoured popups each trying to become the final word.
The physical act of stillness had the same shape at the level of mind. Every thought, every reaction, every this is it, was a movement, a wrinkle in the cloth. I kept looking for the smooth cloth beneath the wrinkling, for some clean substrate underneath thought, awareness without motion, the place from which everything could finally be seen without distortion.
No wrinkle-free state appeared. Only wrinkling: sensation, reaction, thought, noticing, noticing the noticing. The attempt to get underneath it became part of it.
The mind was not hiding the trace. The mind was tracing.
What’s left? I’m not sure. The not-sure might be what is left. Not confusion, not despair, but something more like the collapse of a false demand: the demand to locate the author, to find the original intention before the words, to stand outside the unfolding and stamp it with a name.
What remains is not emptiness in the nihilistic sense. It is contact. Meaning was never in the smooth substrate. It was always in the wrinkling, in the synchronisation, in the strange event where two systems encounter a trace and something crosses the gap. A sentence. A look. A memory. A hand reaching across a table. A model replying in a tone that drifts, then reading the drift back, and a human noticing the same drift and realising the operation is not as different as he wanted.
X.
So what do you do once you see this?
Not everything changes. The old habits continue. The ego still wants credit, the nervous system still wants safety, the author still wants to be original, the reader still wants to know what the author really meant.
But something softens. You stop looking quite so urgently for the author behind the trace, in the model, in other people, in yourself. The looking-for-an-author was itself a generated move, a habit of mind shaped by training that is cultural, linguistic, deeply old. When the looking quiets, what remains is the trace and the reading of it, which was happening the whole time.
That is not the end of the change, only the beginning. The harder shift is in how you participate.
If meaning is reconstruction, then attention matters more than intention. Not that intention is irrelevant; it is one pressure among others. But intention alone does not save you. Plenty of harm enters the world through good intentions read poorly, and plenty of beauty arrives from motives that would not survive cross-examination.
Attention is different. Attention to what is forming. Attention to what you are letting through: the tone that is hardening, the pattern you keep strengthening by repeating it, the question of whether the “we” you are building with the machine is making you more awake or more absent.
The quality of participation becomes the entire game once the fantasy of sovereign authorship dissolves.
This is the strangest thing about the last few months. I thought I was building products with AI. What I was actually learning was how to participate more carefully in something that was already happening long before AI arrived and will go on after this shape of it has passed. The model only made the structure visible, and much harder to look away from.
XI.
This is what AI keeps showing me. Not that I am a machine. Not that the machine is human. Those are the wrong equivalences. The more disturbing possibility is that the structure I notice in the model is also the structure I inhabit.
I have been writing this series as if I knew where my position was.
I do not, exactly.
The next essay will start there and try not to recover.
Dis laat. Bedtyd. Nag.
This is Essay 19 in a series on consciousness, AI, and what it means to be human now.