Writing in the Age of AI: What the Machine Cannot Imitate
In this article
The question arrives regularly now, with varying degrees of anxiety: will AI make writers obsolete? The answer varies by what we mean by “writing” and what we mean by “obsolete” — and unpacking both questions is, I think, more useful than either a dismissive reassurance or a catastrophist alarm.
Let me try to say something that is actually true.
What AI-Generated Text Is
A language model generates text by predicting, at each step, what token (roughly, what word or word piece) is most likely to follow the sequence so far — weighted by the model’s learned sense of what completions are good according to the objective it was trained on.
This process has produced results that are genuinely remarkable. The text is fluent. It is coherent over paragraphs. It follows genre conventions accurately. It can write in many established styles. It can summarize, synthesize, explain, and describe with impressive reliability. These are not trivial capabilities.
But the process has a specific structure: the model generates what is expected of a text of this type, in this context, from this prompt. It optimizes toward the center of the distribution of what good text looks like. It is a remarkably sophisticated averaging device — producing, at high quality, what a large aggregate of human writing would suggest is the appropriate continuation.
This is genuinely useful for a large proportion of writing tasks: documentation, summaries, emails, reports, first drafts of content with fairly specified parameters. These tasks call for competent, genre-appropriate text, and AI produces it reliably.
The Center of the Distribution
What the averaging process does not produce well — and this is structural, not a capability gap that will be closed by the next model version — is the text that earns its value precisely by departing from expectation.
Consider what makes a sentence memorable. Almost never is it that the sentence does what the preceding context predicted. The memorable sentence arrives at something true in a way that had not quite been found before. It is not a surprise for novelty’s sake; it is a surprise that, on reflection, feels like recognition — yes, that is what it is, stated precisely in a way I couldn’t have anticipated but recognize as right.
This is what I mean by voice. Not style in the sense of word choice patterns or sentence length distribution — those are surface features that can be imitated. Voice is the accumulation of an individual’s particular relationship to language: what they notice, what they find worth saying, what connections they make between disparate things, where they choose to stop rather than elaborate, where they elaborate rather than stop.
Voice carries a person. It carries their specific experience of existing, their particular angle on the phenomena they write about, their specific obsessions and hesitations and pleasures. None of this is in the weight matrix. It is in the person.
The Argument from Experience
Consider what distinguishes writing that lasts from writing that does not.
It is almost never pure technical proficiency. Many technically proficient writers are forgotten. What endures tends to be writing where a particular consciousness has pressed itself into language so specifically that reading it is, in some sense, an encounter with a person who has seen something particular and found a way to transmit the seeing.
George Orwell did not write memorable political prose because he had a sophisticated model of what good political prose sounded like. He wrote it because he had specific experiences — fighting in Spain, poverty in London and Paris, working in a colonial bureaucracy in Burma — and a specific moral drive that pressed on him to say certain things that needed to be said, in a way that came from who he was.
Joan Didion’s prose has the quality it has because she was that person in that body with that particular relationship to anxiety and California and the specific disintegration of the 1960s. The prose is inseparable from the perceiver.
An AI model can produce Orwell-like or Didion-like text. It is recognizable as an imitation. It is accomplished. It is missing the thing that made the originals matter: the specificity of a real consciousness encountering real phenomena and trying, with the available tools of language, to transmit something true.
What This Means for Writers
This is not an argument for writer complacency. Several things follow.
The work that doesn’t require voice will be automated. Much of the writing work that exists in the world — corporate content, templated reports, standard documentation, SEO-oriented articles, genre fiction following established formulas — does not particularly require the specific contribution of a specific person. These tasks will be performed increasingly by AI, and this is a genuine economic displacement for people who have earned their living doing them.
The writing that requires voice becomes more valuable. As competent, generic text becomes abundant and cheap, writing that carries specific human authority becomes rarer relative to need. There is more demand than supply for writing that has genuine perspective, earned through real experience and genuine thinking, and expressed in a voice that is specifically someone’s rather than statistically plausible.
The first draft is no longer the hard part. For writers who were slowed by the mechanical work of getting words on the page, AI assistance genuinely helps. Getting a passable structure up quickly, generating options to select from and improve, unsticking at the sentence level — these are legitimate uses. The question is what you bring to the revision that AI brings to the draft.
Knowing what you think remains the hard part. The limiting factor in most writing is not generating text. It is knowing what you actually believe, what you actually experienced, what you actually want to say — and having the intellectual honesty and clarity to say it rather than saying what the form demands or what the reader might prefer. This was always hard. AI makes the first stage easier and makes the second stage more explicitly where the value is.
The Responsibility
There is a responsibility that comes with this view, and it is uncomfortable.
If what distinguishes meaningful writing from competent AI output is specificity of experience and honesty of thought, then the obligation for anyone who wants to write meaningfully is to actually have the experience, actually do the thinking, and actually be honest about what they find there.
This sounds obvious and is in practice rare. Most writing — before AI, not just because of AI — occupies a middle space: technically competent, structurally sound, saying things that are true enough, expressing views that are appropriate for the form and context. Not dishonest. Not hollow. But not specifically anyone’s either. Not reporting something that could only have been reported by someone who had looked at this particular thing and thought about it in this particular way.
AI will produce that middle space text more reliably, at higher volume, at lower cost, than most human writers. If that is the competitive space, the competition is already over.
The space AI cannot occupy — not due to its limitations today but due to what it is — is the space of writing that is specifically someone’s. Writing that carries the irreducible fact of a consciousness having encountered the world and tried to say something true about what was found there.
That is a high standard. Most human writing doesn’t meet it either. But it is the standard that makes writing matter — and it is wholly, structurally, permanently on the human side of whatever line AI has drawn.
The machine is averaging. You are not. That distinction only counts for something if you use it.