Essay
The disappearing middle of writing
What AI is doing to the texture of professional prose.
There is a quiet unease that has begun to settle, in the last eighteen months, over anyone whose work is to read serious writing for a living. It has not yet had its name. We expect that it will be the central editorial feeling of the next decade, and that it will outlive several of the louder feelings currently in fashion.
Begin in the offices of a quarterly review of essays and long reportage, on a higher floor of a building an older industry vacated some years ago. The editor is at her desk after seven. She has been at this magazine, or magazines like it, for twenty-two years. Her readers know her ear; her contributors, depending on the verdict, fear it. There is a stack of pieces in front of her — five long submissions under consideration for the spring issue, three by staff writers she has read for nearly a decade, two by outsiders the magazine is auditioning. She begins with the staff piece on the top of the stack.
It is a profile of a man who runs a chip foundry in central Texas. The writer is one of her most reliable; she has been editing him since he was a fact-checker. The piece is lucid. The piece is reported. The opening, she thinks, is good — three short sentences, a vivid scene at the foundry’s gate, a pivot into a longer sentence that lands the stakes. By page three, she finds her pen has not moved. That is unusual, in her work, and not in the way she likes.
She sets it down and picks up the second. It is from a younger staff writer, a woman whose first feature ran in the autumn issue and was well-received. The piece is about a strike at a logistics warehouse outside Memphis. The reporting is generous; the structure is sturdy. The opening is also three short sentences, then a longer one. The pivot on page two — what is striking, in the strikers’ own framing, is — is a construction she has read very recently, somewhere, and not in this writer’s voice.
The third piece, by a young historian on contract, is a review essay on a recent book about Cold War science policy. The fourth, by an outsider, is a reported piece on hospice care. The fifth, by her most senior staff writer — a man whose sentences she could once have identified blind from a stack of fifty — is an essay on a small museum in upstate New York. She works through them at the pace she has worked through such stacks for two decades. By the time she finishes the last, she is bothered, and she is not yet able to say why.
She gets up. She makes coffee in the office kitchen, slowly, because moving slowly is what she does when something is wrong and she does not yet know what. She comes back to her desk, sits down, and instead of returning to the stack, she pulls up the half-finished draft of her own editor’s letter for the issue. She reads its opening paragraph.
She sets it down very carefully. She reads it again.
The three short sentences and the longer fourth. The what is striking pivot. The em-dash deployed where a comma would have done the same work and asked for less attention. The closing that gestures rather than commits. They are all there. They are in the five pieces on her desk, in five distinct hands, and they are in hers.
She does not, yet, know what to do about it. She does not know how to write a note to a writer she has been editing for nine years that says your prose has begun to sound like four other writers’ prose, and also like mine, and I cannot tell whether the problem is you, or me, or the tools all of us have been using, or whether what I am calling the problem is even a problem at all. She does not know whether the writers know. She does not know whether, if she told them, they could stop. She is not certain, sitting at her desk with the coffee gone cold, that she could stop.
The unease does not have a name yet. But the question that follows her home tonight is the question that will follow many of us, regardless of what we read or what we write: what, exactly, were we reading for?
We were not reading for information. Or rather: we were reading for it, and we were also reading for the other thing, the one that almost no editor, almost no hiring committee, almost no peer reviewer, ever quite said out loud.
That was the trick of the manuscript. It made it look as though information was what we were buying. It was not. Information was the unit; the differential signal carried in the information — the texture that told us whose information, and what-the-writer-thinks-she-knows, and how-far-she-is-from-the-place-where-she-would-stop — was the product. The sentence was a proxy, and like every proxy, it worked until something arrived that broke the relationship between the proxy and the thing.
The five manuscripts on our editor’s desk, in our story, broke the proxy. They contained, by any reasonable measure, the information her readers had been promised. The strike was reported. The foundry CEO was profiled. The Cold War book was reviewed. By the standard of whether the writing happened — yes, it happened, and it happened faster, and at lower cost, than any comparable stack of submissions five years ago.
But notice what the editor did when she sat with the stack. She did not, in the end, mark up the prose. She sat there because she was registering, without yet being able to articulate it, three things the writers were no longer carrying on the page: which of the staff voices she had been editing for years was now the one whose ear had loosened, and how she would say that to him; which of her younger writers had been ambitious enough to fight the tool’s pull toward smoothing and which had not; and what her own editor’s letter, in its current draft, had unwittingly conceded.
That is what the editor was reading for. She was not reading for the information. She was being trusted, by a magazine that paid her to be trusted, to register the things in a manuscript that the manuscript was not, exactly, saying — to know which of these writers was on a sentence-level upswing and which was tired, which piece would survive a year on the shelf and which would read as dated by August, which sentence in any given paragraph had been thought toward and which had been arrived at by the soft current of habit. Voice was never decoration. Voice was data.
This is the distinction that AI is now making impossible to hide.
For most of the history of professional reading, the information and the signal were welded together. They could not be separated by the writer, and they could not, in any case worth caring about, be separated by the reader. A clean paragraph on a complex subject required someone who had thought clearly about that subject. The fact of the clean paragraph was evidence about the person who produced it. The text and the trace were the same artifact, because the only way to produce the text was to make the trace.
That is no longer true. The middle now ships smoothed. A young historian’s review essay, an analyst’s investment memo, a doctor’s case note, a lawyer’s letter to opposing counsel, a journalist’s profile of a foundry CEO — these were the genres in which voice and rigor had reliably travelled together, and these are the genres an off-the-shelf model now produces with reading-grade competence in seconds. The information arrives, formatted, clean, and stripped of the friction that used to be how the reader could tell what kind of mind had touched the page.
What is left of voice, after the smoothing, is not gone. It is redistributed, and the redistribution is the move that matters. At the high end — the writers whose voices are the entire reason we read them, the Robinsons and the Didion-imitators-who-actually-managed-Didion, the historians whose paragraphs sound like nobody else’s — the smoothing tool produces nothing usable. The model can write like such a writer in the imitative sense that a clever twenty-four-year-old can write like her, which is to say not really, and not for a reader who knows what they are looking at. At the low end — the text message, the meeting recap, the customer-service email, the polite end-of-week update — voice was never a load-bearing signal. Nobody was reading those for who-wrote-them; the smoothing tool produces results that are, if anything, slightly better.
It is the middle that goes. And the middle, in the way professions and institutions have used the word for the last hundred years, was where almost everything happened.
Here is the test. Take a paragraph from a serious magazine, published in 2018, by a writer of the second rank — competent, working, employed, not famous. Give an off-the-shelf model the topic, the rough thesis, the audience, and the genre. Ask for a paragraph. Read both. The model’s paragraph is publishable. By the standard of would a non-expert reader notice a difference — no, they would not. By the standard of would the writer’s own editor have known it was not him — yes, instantly, because her relationship with that writer’s sentences was the working capital of her judgment, and the model’s paragraph does not carry the working capital. By the standard of can the model produce, on demand, a paragraph indistinguishable in voice from this specific writer — also no, and that “also no” is the part of the test that has been most badly misreported. The model does not write like anybody. It writes from the average of everybody, and the average is what the middle now sounds like.
That is the entire game. The middle of professional writing — the band between the unmistakable voices at the top and the disposable prose at the bottom — has been compressed onto a smoothed median, and the compression is happening fastest in exactly the band where the labor of writing was doing the most invisible institutional work. Voice in the middle was how a hiring committee told a working journalist from a posing one. It was how a senior partner told an associate who had been thinking from one who had been reading off the model. It was how a peer reviewer caught the second-year graduate student who had hidden a shaky chain of reasoning under presentable sentences. It was how an editor — like ours, at her desk — distinguished the writer she would back from the writer she would let go.
This rearranges who is exposed and how. The writer at the top — the unmistakable one — is, at the end of our editor’s long evening, more valuable than she was the year before, not less. Her voice has not depreciated; the median surrounding it has. She is more identifiable than ever, because almost everyone around her now sounds the same. The writer at the bottom is unaffected, because nobody was paying her for voice in the first place. The writer in the middle — and most working writers, in any given decade of any given profession, are in the middle — is in a different position. She can produce middle-grade prose at the new speed, but the prose she can produce is no longer what she was being read for, and the prose she was being read for, increasingly, she does not get to produce on her own often enough to remain the writer she was.
The question we were left with at the end of the last section was what, exactly, were we reading for? Here is the answer. We were reading for the trace — the part of the prose that told us, sentence by sentence, what kind of mind had touched the page — and we were reading for it because the trace was, in every institution that had a serious editorial culture, the audit instrument the institution could not otherwise afford. The information was the unit. The trace was the product. When we can still find the trace, the instrument multiplies the writer. When the trace has been smoothed, the instrument exposes the institution that was reading for it. The middle was always the band that did the institutional work. The band is just thinning.
It is tempting, having looked at our editor at her desk, to file the unbundling under “aesthetic loss” and move on. We are not going to do that, because the same machinery is now visible in places where the stakes are higher than how a magazine reads.
Aesthetic loss is the easy case. When voice compresses in magazine prose, a profession of editors mourns a register, fees on freelance work soften, and the literary-fiction shelves continue to do what they have always done. Painful in the way that any compression of a tradition is painful, but contained. The harder cases are the ones where voice was carrying, all along, the part of the work that audits whether the work is right. We want to look at two of those.
The first is science, and we will use the peer-review system as the example, because peer review is where the smoothing of professional prose has left the cleanest published evidence. Researchers tracking the frequency of specific words across the PubMed corpus have found, since the public release of large language models in late 2022, sharp and sustained excess use of a small set of words — delve, intricate, showcasing, meticulous, commendable — at a scale unmatched in the corpus’s prior history.[1] This is the surface tell, and it is, in itself, only a curiosity. What it indexes is the deeper phenomenon: the median scientific manuscript now arrives, at the reviewer’s desk, in a register the model finds most probable, which is to say, in the register that has the least friction left.
The labor of writing the manuscript compresses. The wrapper detaches. So far, the story rhymes with the editor’s.
Then it diverges, because of what happens to the reviewer.
The reviewer’s job, of which a remarkable amount of professional knowledge work is in fact composed, was to read a manuscript not for its surface but for its tells. The good reviewer caught the figure that did not, on close inspection, support the claim it accompanied. She caught the literature review that had been smoothed over an inconvenient prior paper. She caught the sentence in the discussion that hedged a result by half a degree more than the data could bear. What she was reading for, in every case, was a discrepancy between what the prose sounded like and what the underlying work actually demonstrated. The prose, until very recently, had carried the trace. A manuscript that read like a hurried first draft warned the reviewer to look at the figures twice. A manuscript that read like the kind of paper a careful lab produces told her she could give the figures a single, attentive pass.
That instrument is failing. When the prose arrives smoothed — when every manuscript, regardless of who actually did the work, reads in the same competent middle register — the reviewer’s pre-attention discrimination has nothing left to discriminate on. She has to read every paper the way she used to read only the ones that had alerted her. The instrument that did the routing was the prose; the smoothing has broken the router. A growing body of work, including a Stanford effort to estimate, at scale, the share of conference reviews and submitted papers themselves that have been at least partially model-written, suggests the effect is real, measurable, and accelerating.[2]
This is the epistemic version of what the previous section described. There, we said voice was the audit instrument. In science, the audit instrument required by peer review — the reader’s pre-attention triage on prose — is the same audit instrument that the smoothing erases. The wrapper and the contents were not welded together by accident. The daily exercise of reading for the tell was what kept the literature trustworthy enough to be worth reading. A field that hands the production of its prose to AI without asking how its reviewers will stay reviewers is a field that is, quietly, importing into its archive the claims its own triage system would once have caught.
This is not an argument against the tools. The tools are real, they help non-native English speakers participate at a higher rate, and they accelerate the unpleasant cleanup work of writing up. It is an argument for redesigning what peer review is asked to do, what it is given the time to do, and what the audit trail of a published paper actually looks like. The premise of the last seventy years of scientific publishing was that how a paper read was downstream of how the paper had been done. That premise is gone. Whatever replaces it has to do, deliberately, what attention-on-prose used to do by accident.
The second case is judicial opinion-writing, and we are aware that this sounds, on its face, like a smaller problem than science. It is not. It sits in a place fewer people look.
Common law is, more than any other institution we have, a reading practice. A working judicial system in this tradition is a system in which judges and clerks read other judges’ opinions, decades and centuries deep, and reason forward by analogy from how those judges had reasoned. The reasoning leaves its trace in the writing. The writing — its voice, its texture, the moves it does and does not make — is how a later judge tells the case that genuinely controls from the case that merely cites well. Whose reasoning, in what register, with what reservations, against what background of contrary authority are the questions a serious reader of an opinion is asking by the second page, not the tenth.
Most of the public conversation about AI in law has focused on the cheap version of the problem: lawyers who submit briefs containing cases that do not exist, of which the New York sanctions in Mata v. Avianca are now the canonical example.[3] Hallucination is real, and it is, on the timescale of this essay, the part of the problem that will be solved. The case-citation tools that produce the briefs are already, in 2026, far less likely to hallucinate than they were in 2023, and the firms are already, by policy and by software, catching what slips through. The cheap version of the problem will, over the next half-decade, mostly fix itself.
What will not fix itself is the smoothing. When clerks draft opinions from model output that they then revise toward the judge’s voice — and when, increasingly, the judge’s voice itself has been formed inside the same set of tools — the system of common-law reasoning loses the texture that did its load-bearing work. Two opinions that should sound nothing alike, because they are doing different kinds of doctrinal work for different kinds of cases, begin to sound the same. A later judge reading those opinions for guidance finds them, on the surface, equally usable, and is no longer being given the prose-level signals that used to tell her which opinion was carefully worked through and which was assembled from the available median of yesterday’s reasoning.
The cost is not visible in any single opinion. The cost accumulates across decades, in an institution that, more than most, compounds by texture. The substrate of legal life is not statute; statute is the surface. The substrate is the long body of opinions that judges and lawyers and law students have been reading and re-reading for two hundred years, and the readability of that substrate — its capacity to teach the next generation what good reasoning looks like — depends on the substrate not having been smoothed.
The editor’s story, at the start of this essay, was never really about a magazine. It was about an institution that read the middle of professional prose for differential signal, and an instrument that arrived which could produce the middle without producing the signal. That story is now being told, simultaneously, in a peer reviewer’s inbox, in a judicial clerk’s first draft, in a hiring committee’s writing-sample folder, in a graduate seminar’s discussion of two students’ essays that read uncomfortably alike. The aesthetic loss is real. It is also the least important part.
By now the question is no longer whether the smoothing is happening. The question is what people are doing about it, and what the cost will be when it turns out that none of what they are doing is enough.
The first response is refusal. Through 2023 and into 2024, the dominant institutional posture across publishing, scientific journals, and the courts was containment. Nature and Science published author guidelines in January 2023 stating that large language models could not be listed as authors and that any use had to be disclosed in methods; the major society and university-press journals followed within months.[4] The New York Times, in early-2024 internal guidelines reported in trade press, instructed its newsrooms that staff were not to draft published copy with generative AI and required disclosure for any use. Federal judges, beginning with Brantley Starr in the Northern District of Texas in May 2023, issued standing orders requiring counsel to certify either that no portion of a filing was drafted by AI, or that any such portion had been verified by a human reviewer.[5] These were not, on their face, denials that the tools mattered. They were, on their face, sober responses to instruments whose epistemic boundaries no editor and no court had yet figured out how to draw. But functionally, the bans and the disclosure mandates served as reasons not to think hard about the deeper question, which is what the instrument, properly bounded, was already doing to the texture of the work itself. Disclosure is not a redesign; it is the institutional act of asking a writer to flag what the institution is no longer in a position to detect on its own. Refusal bought time. Time was used poorly.
The second response is substitution: the assumption that, since AI compresses the labor of writing, an organization can shed writers in proportion to the labor compression. The canonical public case is the Sports Illustrated byline scandal.
In November 2023, Futurism reported that the operator publishing Sports Illustrated under license had been running articles under bylines, with author photographs, of writers who did not exist — the names were AI-generated, the photographs were generated by a separate AI service, and the articles themselves had been at least partially model-written.[6] Within forty-eight hours of the report, the magazine’s parent took the articles down and ended its contract with the third-party content operator responsible. The story is now read, sometimes, as the cheap version of the AI-substitution problem — fake bylines, fake photographs, a magazine of record embarrassed.
The lesson is not the embarrassment. The lesson is what the substitution had been doing in the months before the report. Sports Illustrated did not, presumably, set out to deceive its readers; it had outsourced a category of routine sports coverage to an operator whose business model was producing the middle of professional prose at a fraction of the cost of producing it with people. The information was the unit. The trace had been removed by design. What broke the arrangement was not, in the first instance, the readers — most readers had not noticed — but a journalist whose business it was to read the middle for the trace, who found it absent, and who tracked the absence back to its source. The headcount had been reduced. The work had not. The work had simply moved to the reader, who is unpaid, untrained, and free to leave.
CNET’s earlier experience with model-generated finance explainers, retracted and corrected at scale through January 2023, is the lower-profile prototype.[7] The pattern is the same: substitution as a strategy did not produce work that was wrong in the way a careless reporter is wrong; it produced work that was clean, smoothed, presentable, and could not be trusted, because the institution that was supposed to be reading it for the trace had been the one that designed the trace out.
The third response is augmentation. Augmentation is the most thoughtful of the three, and it is the response most major newsrooms, the most established scientific publishers, and the largest professional law firms have publicly adopted. The Associated Press is the canonical example: the AP has run algorithmic earnings reports since 2014 under an “augmented journalism” framing, in which the routine work of producing quarterly-earnings summaries is handed to software, with human reporters redirected to the stories the algorithm cannot do.[8] Reuters, the Washington Post, and Bloomberg operate analogous pipelines for sports recaps, real-estate transactions, and corporate earnings; in 2024 and 2025, The Guardian, The Atlantic, and The New York Times publicly described internal AI tooling for research assistance, headline generation, transcription, and draft-stage suggestion, in every case framed as multiplying senior judgment rather than replacing it. The framing is consistent across institutions: AI is a tool that multiplies editorial taste; it is not a substitute for it; the senior reporter does not become obsolete, the senior reporter becomes more leveraged.
This framing is correct as far as it goes, and it is significantly better than the first two responses. It is also consonant with the argument we made in section two — that voice was the audit instrument, and that an instrument which does not get tired multiplies the value of editorial judgment that knows what to read for. Augmentation, done seriously, is the response a serious publication ought to be running at.
But augmentation, as currently practiced, has a hole in it that none of its proponents has yet filled. The hole is the one we named in the science section: if the daily exercise of reading for the trace was what kept the trace recognizable, and if the new arrangement hands the production of the middle to the instrument, then the question of how the next generation of senior editorial judgment is formed — how today’s juniors become tomorrow’s editors, reporters, reviewers, and judges — is suddenly an open one, and it is being answered, in most institutions, by improvisation. The augmentation-first newsroom has its senior editors directing the tool more effectively. It does not yet have an answer for what its young reporters will know in fifteen years that today’s senior reporters learned by writing the work the tool now does.
That is the question we will spend the next section on. None of the three responses has reached it. Refusal does not see it. Substitution forecloses it. Augmentation gestures toward it and then, mostly, hands it back to the human-resources department. It is the question that decides whether the professions that depend on the middle of writing — journalism, science, law, and the editorial culture of the magazine itself — exist on the other side of this transition in any recognizable form. We think it is the most important question this essay has named.
We have arrived, finally, at the question that will decide whether the professions of serious writing exist on the other side of this transition: how does the next generation of voice get formed when juniors no longer produce the middle of professional prose under their own power?
The honest answer is that no one yet knows. We are going to spend this section laying out why — and what is at stake in the not-knowing — and we are not going to pretend, at the end of it, that we have solved the problem. The honesty is the section’s argument. Anyone who claims to have figured this out in 2026 is selling something.
Begin with what the apprenticeship system actually was, for the trades of serious writing. A young journalist began on the desk, writing weather copy and police-blotter rewrites and the small-town meeting recap — the most boring work the paper produced. The work was bad. The desk editor marked it up in pencil, the next day’s work was less bad, the year’s work began to develop a recognizable cast of habits, and by year seven there was a voice on the page that was unmistakably this reporter’s and nobody else’s. The young scientist began with the methods section nobody senior wanted to write; the young lawyer began with the brief on the four-hundred-page record nobody senior had time to read; the young magazine editor began at the proofreader’s desk, then the fact-checker’s chair, then the section editor’s queue, then — much later — her own column. The mechanism was not the formal training, which existed but was always lighter than its institutional reputation suggested. The mechanism was the labor. The seniors were made by ten thousand corrected drafts, and the drafts were the work the institution was being paid to produce anyway. The reproduction system and the revenue system were the same activity, performed in the same chair.
The new arrangement breaks this almost completely. The young journalist drafts the weather copy by prompting the tool, accepts the output, and posts it. The young scientist generates the methods paragraph by feeding the model her experimental notes and lightly editing the result. The young lawyer drafts the brief by asking the tool to summarize the record, then reorganizing the model’s organization. In each case the junior is still working — she may, in raw hours, even be working harder — but she is doing fundamentally different work. She is prompting and revising rather than producing and being corrected. Two activities that feel adjacent. They are not.
Prompting requires a model of what good output looks like, and that model is built almost exclusively by producing bad output, having it corrected, and producing less bad output next time. The prompter without the production history is the inspector who has never welded, the editor who has never written. She can pattern-match what looks bad on the surface. She cannot reliably catch what is quietly bad — the paragraph that reads correctly except for the one phrase the model has smuggled in from the average of everybody, and which the careful writer would have killed before the sentence was finished.
This is not a sentimental claim. The literature on craft transmission, which the sociologist Richard Sennett synthesized as well as anyone in the last decade, treats the long apprenticeship as essentially non-substitutable for the kind of judgment senior practitioners actually use.[9] You can teach the explicit lessons in seminars; the implicit lessons — the ear, the sense of what this writer’s first instinct should be at this point in this paragraph — require thousands of repetitions of the producing-and-being-corrected loop. The loop, in writing, was the writing.
What is being tried, in newsrooms and journals and law schools that are taking the problem seriously, falls into roughly three buckets. The first is apprenticeship by audit: structured programs in which juniors edit AI-drafted output under senior supervision, with the senior using the review session itself as the teaching moment. This is a sensible response, but it has a scaling problem and a depth problem. The scaling problem is that senior time is the bottleneck the institution was trying to alleviate by deploying the AI in the first place. The depth problem is that auditing is a thinner slice of the original apprenticeship — the junior catches some lessons by reviewing, but loses the lessons that came from the felt experience of producing the wrong sentence herself, and of having to find the right one without any draft on the page beneath her hand.
The second bucket is simulated production: making juniors write under AI-off conditions — in proctored rooms, in blue books, in longhand — so that they retain the ability to produce. Several major American journalism and law schools have moved a larger share of their written assessment back into proctored AI-off settings since 2023; The New Yorker’s long-running fact-checking apprenticeship has historically functioned as a forced longhand-equivalent immersion in voice; several federal judicial clerkships are reported, anecdotally, to have begun running clerks through periods of AI-prohibited drafting as a deliberate developmental exercise. The early results are interesting and, importantly, partial. They suggest that simulation can keep some skills alive but cannot fully reproduce the emergent voice that used to come from doing the actual work, for real readers, in real institutions, repeatedly. Simulators teach pilots how to fly. They do not teach pilots how to be the kind of person who knows, on instinct, when something is about to go wrong.
The third bucket, and the most candid, is waiting. Many institutions have decided, implicitly or explicitly, that the current cohort of senior writers, editors, reviewers, and judges is sufficient for the next decade and that the question of generational reproduction can be addressed later. This is an institutional bet that the problem will resolve itself — that better tools, or better training programs, or simply better juniors, will materialize. It is not a strategy. It is the absence of one.
We are not in a position, in 2026, to say which of these will work, because none has been running long enough to find out. What we can say with some confidence is that none of them is, on its own, sufficient — and that the cost of getting this wrong is not borne in the year the bet is placed. It is borne fifteen years later, when the cohort that should have become senior editors and senior reporters and senior reviewers and senior judges shows up underweight on the very ear, instinct, and voice those titles imply.
That, by itself, would be enough to call this the most important question in the essay. But the formation mechanism is, in fact, the smaller of the two questions an institution-of-serious-writing faces under these conditions.
The larger question is what a profession of writing is, when its reproduction mechanism breaks.
A profession, in the strong sense, is not a job category. It is an institution that controls the production of its own producers — that decides who gets credentialed, what knowledge counts, what conduct gets a member expelled, and what the public is owed in exchange for that closure.[10] Journalism, in this sense, is a profession; so is the academic discipline that runs its own peer-reviewed journals; so is the bar. Each has, until very recently, been able to claim that its members produce a kind of prose that distinguishes them from people outside the profession — a kind of prose that is, in fact, evidence of a kind of training. The labor market is downstream of that claim. When the claim weakens, the labor market does not immediately collapse — the existing senior writers continue practicing, the existing credentialing bodies continue credentialing — but the institution slowly loses what justified its claim to special standing in the first place. It loses the line between the people who do this and everyone else who can prompt a model.
That line is being attacked from below as well. The writer Maggie Appleton’s barefoot developers — the coming class of end users who will produce the software they need without becoming software engineers in the institutional sense — extends naturally to writing.[11] The barefoot reporter covers her own city-council meeting using a transcript and a model. The barefoot reviewer writes her own book review for the local paper that no longer employs a critic. The barefoot legal writer drafts her own demand letter, her own response to opposing counsel, her own appellate brief on a small matter. None of these barefoot practitioners is replacing the senior writer at the top of the profession. All of them are, however, redrawing the edge of the profession’s claim to its labor market.
What is happening, in other words, is a double erosion. The credentialed pipeline is unable to reproduce its senior writers as fast as the senior writers retire. The lay public is, simultaneously, encroaching on the simpler end of the work. The professions of writing are being thinned in the middle and frayed at the edges at the same time. A profession can survive one of these. We are not certain any profession of serious writing can survive both, in their current institutional form, at the speed at which they are now happening.
This is the question we want to close on. Not “will the professions adapt?” — every institution under pressure adapts, in the trivial sense that something will exist where they used to be. The question is whether what comes out the other end is recognizable as a writing profession at all, or whether the institution is replaced by something else: a credentialed thin layer of accountability over a substrate of barefoot practice, a small guild reduced to the unmistakable voices at the top, a regulated infrastructure for the disclosure of model-assistance. The answer matters because the institution is what produced the people we have been writing about, including the editor at the start of this essay, and what replaces it will produce — or fail to produce — the editors and reporters and reviewers and judges who will be reading and writing in 2050.
The editor, by now, has gone home. Tomorrow she will go back to her office; on Friday she will send notes to five writers and the editor’s letter to layout; in fifteen years she will retire. The question is who will be at her desk.
Friday will look ordinary. She will accept three of the five pieces for the spring issue and ask for revisions on the other two. She will catch, in the staff profile of the chip foundry CEO, a paragraph on page seven that nobody else would have caught — a paragraph that says, in the smoothest possible English, exactly the thing the foundry’s communications department most wants said about its founder, and that, with a single small movement of the writer’s pen, can be made to say what the reporter actually believes. She will write a two-paragraph note to her senior staff writer that does not mention the smoothing directly but that he will, if she has chosen her sentences correctly, recognize. The issue will close. The magazine will go to press. The writers will go home pleased with themselves, mostly rightly so — they are good at this, they will get better, they will stay at the magazine. Nothing on Friday will indicate, to anyone watching the magazine from the outside, that anything has changed.
Everything has changed. The editor has spent her week doing the only part of her job that has not yet been smoothed — the part where she knew, in a paragraph, the thing that mattered, and the part where she could say it to a writer she had been editing for nine years in a sentence that did not need to be explicit. That part is small. It is also what the magazine is, increasingly, paying her for. The young staff writer has spent her week doing the part of her job that the previous generation of senior writers did at her age, except that the part she is doing is no longer the part that was forging them. She is producing more, and forming less, and she does not yet know that this is the trade.
Somewhere in a university town across the country, a peer reviewer of comparable seniority is having an analogous Friday. She has read four manuscripts before lunch. Three of them arrived in the smoothed register that has become the working baseline of her field. She caught, in the fourth, a passage of methods that did not, on inspection, support the analysis that depended on it. She does not know, and cannot easily find out, whether she would still catch the fourth in five years if her practice continued as it is going — if the manuscripts continued to arrive smoothed, and her own attention continued to be calibrated against the smoothing. She is still the reviewer of record, and her reports are still hers, and the journal she reviews for is, on the whole, more able to publish at scale than the journal that preceded it. None of that resolves the smaller, slower question.
A federal clerk at a court of appeals on the East Coast is, at this moment, finishing the second draft of an opinion in a case that, on its merits, is not interesting and will not be much cited. The first draft was assembled in part from model output. She has revised the prose toward what she takes to be the judge’s voice, which she has been listening to in chambers for ten months and which she has read, in opinions, going back twelve years. Most opinions the court issues are like this one. Most weeks, the long body of common-law reasoning absorbs the opinion without incident, and lawyers cite it sparingly, and the doctrine moves forward. The opinion in which the smoothing has erased a distinction that mattered is the opinion we will read about, much later, in a case that turns on it — and not, yet, in the case we are reading about today.
The thing the editor took home tonight — and the thing the reviewer is taking home tomorrow, and the clerk is taking home next week — is not, exactly, fear. It is older than fear, and quieter, and harder to act on. It is the recognition that the work to which they gave their adult lives turns out to have been two things welded together, and that the welding has come undone, and that the smaller of the two things — the trace they actually built their lives around reading and producing, the trace that justified the long apprenticeships and the slow careers and the late evenings spent learning, sentence by sentence, what voice does — is now visible for what it always was. Some of them will look at it and feel quietly proud. Some of them will look at it and feel exposed. All of them will eventually have to decide what they are going to do about the part of the work they no longer get to do every day, that used to be how they kept the rest of it sharp.
The institutions they have built will decide, over the next fifteen years, whether the answer is found at the level of the magazine, the journal, the chambers, or the school. The formation mechanism is broken in a way that no institution has yet repaired. The boundary of the profession is being redrawn from below in a way no editorial body has yet contested coherently. The cohort that should be replacing today’s senior writers, reviewers, and judges is, for the moment, mostly prompting and revising instead of producing, and the difference will not show up on any season’s roster of new bylines. It will show up the season a piece the editor would have caught goes to press, and a passage in it shifts a reader’s understanding of something it should not have shifted, and a different set of editors begin to understand, for the first time, what was actually being lost.
Friday will look ordinary. Most Fridays, for the next ten years, will look ordinary. The day someone notices that the chair at the desk has been empty, in the sense that matters, for some time, is the day this essay will look like it was written about a problem that had a name.
Someone will be reading the spring issue in 2050. We do not, in 2026, know what they will be able to hear.
Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát, and Jan Lause, “Delving into ChatGPT usage in academic writing through excess vocabulary,” arXiv:2406.07016 (June 2024); subsequently published in Science Advances in 2025. The authors document a step-change in the frequency of a specific set of stylistic markers — most notably delves, showcasing, intricate, underscores, meticulous, and commendable — in PubMed abstracts beginning in 2023, with effect sizes far larger than any observed during the COVID-19 pandemic or any prior shock to the corpus. Their methodology compares against pre-2023 base rates for each word and treats the post-release rise as an excess to be estimated. ↩︎
Weixin Liang et al., “Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews,” presented at ICML 2024 (arXiv:2403.07183). Liang and colleagues estimate that 6.5%–16.9% of text in peer reviews submitted to several major machine-learning conferences after the release of ChatGPT had been substantially modified by a large language model, with the share rising sharply across the review periods studied. Companion work has begun to document analogous patterns in submitted manuscripts themselves; the picture is partial and the methodology is contested, but the trend it indexes is consistent across independent attempts to measure it. ↩︎
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). Steven A. Schwartz and Peter LoDuca of Levidow, Levidow & Oberman, P.C., were sanctioned $5,000 by Judge P. Kevin Castel on 22 June 2023, after submitting a brief containing citations to six judicial decisions that did not exist; the citations had been produced by ChatGPT and presented to the court without verification. The sanctions order and the underlying transcripts have been widely reproduced; the contemporaneous reporting is collected, with documents, by Reuters (https://www.reuters.com/legal/transactional/lawyers-blame-chatgpt-trickery-after-citing-fake-cases-court-2023-06-08/) and the New York Times (https://www.nytimes.com/2023/06/08/nyregion/lawyer-chatgpt-sanctions.html). ↩︎
Nature’s editorial position was announced on 24 January 2023 in “Tools such as ChatGPT threaten transparent science; here are our ground rules for their use” (Nature 613, 612), declaring that large language models could not satisfy authorship criteria and that any use had to be disclosed. Science followed with editor-in-chief H. Holden Thorp’s editorial “ChatGPT is fun, but not an author” (Science 379, 313, 26 January 2023). Springer Nature, Elsevier, Wiley, and the major university presses issued analogous guidelines through the spring and summer of 2023. ↩︎
The standing order is titled “Mandatory Certification Regarding Generative Artificial Intelligence” and was entered by Judge Brantley Starr of the U.S. District Court for the Northern District of Texas on 30 May 2023, requiring counsel to file, with every appearance, a certification that no portion of any filing was drafted by generative artificial intelligence — or that any such language had been checked for accuracy by a human, using a print reporter or traditional legal database. Reporting on the order and its proliferation through other federal chambers is collected by Reuters (https://www.reuters.com/legal/transactional/us-judge-orders-lawyers-sign-ai-pledge-warning-chatbots-make-stuff-up-2023-05-31/) and Law.com. ↩︎
Maggie Harrison Dupré, “Sports Illustrated Published Articles by Fake, AI-Generated Writers,” Futurism, 27 November 2023, https://futurism.com/sports-illustrated-ai-generated-writers. The Arena Group, then the operator publishing Sports Illustrated under license, removed the articles and ended its contract with the third-party content vendor AdVon Commerce within forty-eight hours; AdVon disputed elements of the reporting. The Arena Group made further executive changes in the following weeks; we describe only what the contemporaneous public record supports. ↩︎
CNET, then under Red Ventures, began publishing model-written finance and money explainers under the byline “CNET Money Staff” in November 2022; the practice was reported in January 2023 by Frank Landymore and Maggie Harrison at Futurism (https://futurism.com/cnet-ai-generated-articles) and by The Verge (https://www.theverge.com/2023/1/19/23562966/cnet-ai-written-stories-errors-corrections-red-ventures). CNET subsequently issued corrections to dozens of the published articles; editor-in-chief Connie Guglielmo posted an editor’s note on 25 January 2023 acknowledging the use and addressing the corrections. ↩︎
The Associated Press’s automated quarterly-earnings reports were announced on 30 June 2014 in “A leap forward in quarterly earnings stories,” then by AP’s managing editor for business news Lou Ferrara, in partnership with Automated Insights’ Wordsmith platform; the project is the canonical reference point for the “augmented journalism” framing that has structured the major wire services’ subsequent AI deployment. The framing has been elaborated in the AP’s own subsequent staff guidance and in industry-press treatments at Nieman Lab, the Reuters Institute, and the Columbia Journalism Review. ↩︎
Richard Sennett, The Craftsman (Yale University Press, 2008). Sennett’s argument across the book is that craft is reproduced through the long, repeated, corrected exercise of skill, and that the tacit knowledge characteristic of senior practitioners — what he calls the “intelligent hand” — is unavailable except by that route. The literature on professional apprenticeship that builds on Sennett (and on earlier work by Jean Lave and Etienne Wenger on situated learning) treats this as a structural feature of craft transmission, not a sentimental one. The argument in this section is that serious writing, in the senses we have been using the term, is a craft of this kind. ↩︎
Cf. Andrew Abbott, The System of Professions: An Essay on the Division of Expert Labor (University of Chicago Press, 1988), cited in this issue’s cover essay for a related point. The framing we use here — professions as institutions that control the reproduction of their own producers — is consonant with Abbott’s argument and elaborated in Eliot Freidson, Professionalism: The Third Logic (University of Chicago Press, 2001), which treats professional self-reproduction as the third logic of work alongside markets and bureaucracies. ↩︎
Maggie Appleton, Home-Cooked Software and Barefoot Developers, https://maggieappleton.com/home-cooked-software. Cited also in this issue’s cover essay; we extend the analogy here, as the cover essay extended it, to the professions of writing. ↩︎
§