On authenticity, imperfection, and what AI is actually learning from us
We read differently now. The shift is small enough that many people haven’t noticed it, but once you do, it’s hard to read anything the same way again. We have added, silently and without much announcement, a prior question to the act of reading. Before asking what a text means, we now ask whether a person wrote it.
This is new, at least at this scale. Suspicion about the origins of text used to be reserved for a specific class of documents — press releases, legal boilerplate, corporate communications written in that careful, airless voice that had always suggested a committee. Now the suspicion has expanded to cover everything: an email that’s slightly too well-organized, a film review that moves with unusual precision, an essay that seems to have resolved its own uncertainties just a beat too efficiently. We read the sentence and, before we’ve fully absorbed what it says, we’ve begun examining where it came from.
Something in that deserves a pause. We became suspicious of artificial intelligence. In the process, we became suspicious of language itself.
The project of identifying AI-generated text has developed its own taxonomy of tells. The em dash that arrives with slightly too much confidence. The “not X but Y” construction — balanced, rhetorically clean, carrying its two terms in perfect relation. Certain verbs: delve, resonate, unpack. Sentences that locate their point without any visible uncertainty about whether the point was worth the journey. Metaphors that reach interestingly and fail in a particular way. The list circulates, gets debated, gets revised.
But there is a problem built into this taxonomy that tends not to be examined. Those features didn’t originate with language models. They predate them by decades, sometimes centuries. The em dash is a staple of the literary essayist. The “not X but Y” construction has a rhetorical pedigree that runs through the critical tradition of the last century. Many of the vocabulary items now treated as AI markers — nuanced, tapestry, the whole repertoire of polished approximation — were the professional vocabulary of editors and culture writers who reached for them without irony, because they were, at the time, marks of a certain kind of seriousness. Any writer who came up reading serious criticism absorbed these habits precisely because they signaled belonging to a tradition worth taking seriously.
The forensic project of AI detection has therefore become, whether anyone intended this or not, a forensic project aimed at human writing. To treat the em dash as evidence is to treat the critic who used em dashes before ChatGPT existed as evidence. To flag the balanced rhetorical construction is to flag the very sentences that careful writers of the last fifty years were taught to reach for.
Style was once a writer’s signature. The assumption — articulated most cleanly by Buffon in the eighteenth century, but operative long before and long after — was that the way a person writes is inseparable from who they are. Content could be borrowed, summarized, plagiarized; self could not. The style would give you away.
That assumption has been turned inside out. Style is now evidence — not of identity, but of suspicion. The more distinctly a piece of writing sounds like a sensibility, the more the reader wonders whether that sensibility was inhabited or generated. The features that once proved a person was present now raise questions about whether anyone was present at all.
I find this philosophically strange, in the way that a sentence you’ve said a thousand times can suddenly stop making sense if you say it slowly enough.
Universities have become a concentrated version of this strangeness. The professor who once looked at a student’s writing for the emergence of a distinct voice now reads that same emergence differently. Fluency that would have signaled absorbed learning now reads as a possible flag. Logical structure that moves cleanly from premise to conclusion, without the stumbles and backtracking that mark a mind genuinely working through something — this becomes grounds for hesitation. The educator is now, among other things, a forensic reader. Pattern recognition has become part of pedagogy.
The response, in some cases, has been to run AI-detection tools over submitted work — to use one model to detect another. There is something almost comic about this, and something genuinely melancholy. The institution of education, whose purpose was always to produce a certain kind of developed attention, finds itself deploying the tools of a spy agency.
But the real anxiety in the room isn’t about a paper. It’s about something further out and harder to name. What educators are actually afraid of is not the student who used a language model to write an essay on King Lear. It is the student who arrives at adulthood having always been able to skip the difficulty — who has, in some technical sense, read many hundreds of pages, but who has never been stopped cold by a sentence and forced to remain with their confusion long enough to work through it. Who has never learned that certain kinds of understanding are not available at the speed of a summary.
That fear is not about technology. It’s about attention — about what happens to a mind that is never required to stay uncertain for long. And the specific danger of the current moment is that every ambient signal is telling people that this requirement is optional. The model will interpret. The tool will clarify. The summary exists. The friction is unnecessary.
What no one has figured out yet is how to explain that the friction was never a cost. It was the product.
Here is the development I find most genuinely unsettling — and that I think has been underreported in the wider conversation.
Early language models produced outputs that were, in a specific way, too good. Too consistent. Sentences that were well-formed without exception. Arguments that resolved without remainder. Prose with the quality of a room that has been carefully prepared for visitors: everything in its place, nothing left to chance, no sign that anyone had ever actually lived there. That quality — that absence of mess — was itself the tell. Human writing, even at its most careful, is not like that. It has remainder. It is inconsistent in ways that are neither random nor meaningless. The most memorable sentences are often the ones that arrived somewhere the writer hadn’t predicted, and the uncertainty of the arrival is part of what makes them matter.
The models are learning this. Not by being handed a list of imperfections to simulate, but through sustained exposure to enough human language that they have begun generating something that resembles the texture of human unpredictability. And now the better outputs have something — a studied ambiguity, a deliberate residue of the inexact, a clause that doesn’t close quite when you expect it to — that reads, under ordinary attention, like the trace of a mind genuinely in motion.
What’s disturbing about this is not that it’s technically accomplished. It’s what it implies. AI is not learning to write like a good writer. It is learning to write like the version of imperfect that human readers experience as human. Not thought itself — but the phenomenology of thought in process. Not the experience of working through something — but the impression, rendered with increasing accuracy, that working-through has occurred.
And since we have never had very clean access to the difference between those two things — between the actual experience of thinking and the appearance of that experience — this is a harder problem than a technological one. It is a problem about what we mean when we say a sentence feels alive.
Every writer has a specific relationship to the unresolved places in their own work.
You finish a draft and you know exactly where it fails. The argument that doesn’t quite close. The image that almost works. The rhythm that trips in a place you can’t fix without losing something else. These are the passages you return to, the ones that resist being put to rest. They are, often, the places where the writing is still trying to say something it hasn’t found the language for yet.
AI is unusually capable at precisely this problem. It can identify the word that would be more precise. It can restructure the sentence until the rhythm holds. It can resolve the ambiguity that’s been bothering you for weeks. And for a writer who has been staring at their own blind spots long enough to have lost perspective on them, the availability of that kind of assistance is not nothing.
The question isn’t about whether to use available tools. Writers have always worked with editors, second readers, the whole extended process of revision. The question is what changes in your relationship to the unresolved passage when it can always be resolved on request. Whether the place that isn’t working is a problem to be solved or the location of something you’re still in the process of understanding. Whether sitting with a passage that won’t resolve, without knowing yet whether patience or revision is the correct response, produces something in the writer that the clean output cannot.
In my experience, it does. The unresolved passages are often the ones pointing somewhere — toward an idea still forming, a feeling still looking for its name, a place where the thought is genuinely still in motion. But you can’t always tell from inside the work whether an obstruction is structural or generative. And the tool has no way of making that distinction either. It can only remove the obstruction.
The willingness to leave the obstruction in place long enough to find out what it is — that is not a technical skill. And it is exactly what the tool is optimized to make unnecessary.
There is a question at the bottom of all of this that I have stopped pretending I can answer.
When we describe the greatness of a piece of writing — the kind that stays, that shifts how you see something, that you find yourself returning to at intervals you didn’t choose — we tell a story about what produced it that involves the full weight of a life. The accumulated reading over decades. The specific failures. The attention that only this person, in this time, could have brought to these sentences. We say the work is irreducible because the person is irreducible. The sentence means what it means, and means it in the way it does, because of everything that had to happen before it could be written.
But there is another story, and it gets harder to dismiss. Perhaps what we experience as irreducible depth is, at some level, a pattern that can be learned. Perhaps the sense that a life has passed through a sentence is an effect that can be reproduced with sufficient accuracy to become functionally indistinguishable from the original. Perhaps the features we describe as greatness are formulas internalized so deeply, by both writer and reader, that neither party experiences them as formulas anymore — and if neither experiences them as such, in what sense are they still formulas?
I don’t know how to settle that question. What I notice is that I don’t want it settled in the direction the current evidence tends to point. And that preference is not an argument. It is exactly the kind of wishful resistance that asks, sooner or later, to be tested by the thing it is resisting.
We are not, finally, in an age of judging artificial intelligence. We are in an age of being asked, under pressure, to articulate what we thought writing was for — and finding that the answer is harder to state clearly than the long habit of assuming it mattered had ever required us to acknowledge.
What I come back to is something like this: writing is not, at its most essential level, a way of producing text. It is a way of developing the capacity to think carefully about something for long enough that the thinking changes you. The difficulty of finding the sentence — the weeks a passage might sit unresolved, the revisions that move in the wrong direction before they find the right one — is not a cost the process incurs on the way to its output. It is the process. The writer who has been through it is different from the writer who has not. What they have learned is not available in the sentence itself. It is available only in what the sentence required of them to write.
Whether anything produced without that difficulty is, in any meaningful sense, the same kind of thing — that is the question the next decade will answer, probably before most of us are ready to receive the answer.
What I know is that a sentence is not just the record of a thought. It is the record of having been required to think. And the requirement, it turns out, was never incidental.