No, though it does not leave the work untouched. Editing is at least five distinct jobs stacked together, and only the most mechanical layers are genuinely exposed. The rest depends on knowing whether a claim is true, what the piece is for, and who carries the blame when it is wrong — none of which is a property of the text.
Flow diagram of 5 steps: Stop producing the notes a model already produces, Spend the reclaimed time on verification, Decide what the piece is for before touching a sentence, Take the positions a model will not take, Be explicit about what you are signing off.
The question is close to unanswerable as posed, because the word covers a stack of different activities that happen to share a title. A proofreader comparing a typeset page against the corrected copy, a copy-editor imposing a house style, a line editor rewriting a paragraph so it lands, a structural editor telling somebody the middle third does not work, and a commissioning editor deciding the thing should exist at all are five jobs with different skills, different rates and different degrees of exposure. Asking whether AI replaces editors is close to asking whether it replaces people who work with numbers.
Sorted by exposure, the stack has a clear gradient, and it runs almost exactly opposite to the pay scale — which is worth noticing rather than taking as consolation. The tasks a model handles well are the ones with a determinate answer recoverable from the text alone: whether a word is spelled correctly, whether a comma is doing anything, whether a term means the same thing on page two and page nine. The tasks it handles badly are the ones where the answer lives outside the text, in the world or in the reader.
The bottom of that stack was also substantially automated before language models arrived. Spell-checking absorbed a large share of proofreading decades ago, and grammar and style checkers took another slice through the 2010s. The current shift continues that line rather than breaking it, and the previous rounds did not eliminate the job — they raised the floor of what counted as editorial work and moved the boundary upward. The reasonable expectation is another move of the boundary, not an evaporation.
The editorial stack, and how exposed each layer is
| Layer | What it decides | Exposure to automation |
|---|---|---|
| Proofreading | Spelling, typography and consistency of the final artefact against the corrected copy | High — largely automated already, and models close most of what was left |
| Copy-editing | House style, grammar, internal consistency, clarity at sentence level | High for the mechanical half; low for the half that requires knowing what the author meant |
| Line editing | Whether a paragraph works — rhythm, emphasis, register, what to cut | Medium — a model produces plausible suggestions and cannot tell which of them the piece can afford |
| Structural editing | Whether the argument holds, what the piece is for, what has to be rebuilt | Low — depends on the subject, the intended readership and what the writer was attempting |
| Fact-checking | Whether the claims are true outside the document | Very low — needs sources the model does not have and cannot be relied on to have |
| Commissioning | Whether the thing should exist, who should write it, what it is worth | Very low — a judgement about the world, and a decision somebody has to own |
Most arguments about this are conducted as forecasts about capability: the model cannot do a particular thing yet, and the reply is that it could not do some other thing a year ago either. That exchange never terminates, and it distracts from the part of the job that does not turn on capability at all. An editor is a person who is answerable for what gets published. When a legal claim arrives, when a source says they were misrepresented, when a correction has to be written, somebody has to have decided — and a decision requires a decider who can be wrong in a way that costs them something.
This is not a mystical property, it is an institutional one, and it is visible in how publishing is arranged: the sign-off, the named responsible editor, the indemnity, the corrections column. None of those describe a skill. They describe where responsibility has been placed, and a system that produces excellent notes while being incapable of carrying responsibility for them does not fill that slot however good the notes become. The practical consequence is that even heavily automated editing needs a person attached to it, and once a person is attached you have an editor with a changed workload rather than no editor.
The second durable gap is that a model has no access to whether a thing is true. It reads a fabricated quotation and a real one identically, because on the page they are identical and the difference sits outside the document entirely. Editors of any seriousness spend a large share of their attention on exactly this — the figure that is four years out of date, the attribution that has drifted through three retellings, the anecdote the writer half-remembers. That work does not get easier when drafting gets cheaper. It gets more necessary, because the volume of confidently phrased unverified material goes up.
The standard reassurance is that a person will review the output, and as an organisational design that phrase conceals the problem it claims to solve. Reviewing fluent, plausible, mostly-correct text is a harder task than reviewing a rough human draft, because there are no surface signals marking where the errors are. The reviewer has to check everything at the same rate, gets no help from the prose about where to look, and is commonly given less time on the grounds that the draft arrived already polished. A loop arranged that way produces attention fatigue and misses things a slower process would have caught.
Stopping at reassurance would be dishonest, because there is a real displacement and it has a shape. It falls on the bottom of the stack, on piecework, and on the entry rungs — and those are the same rungs the profession has always used to train people. A publisher that no longer pays for a first proofread has not only removed a cost, it has removed the job that taught somebody to read at that level of attention. The senior editors who cannot be replaced were all produced by the junior work now being removed, which is a slower and far less visible problem than the immediate loss of the roles.
The second shift is a new task disguised as an old one: cleaning up machine-generated drafts. It is commissioned as copy-editing because the text looks clean, and it demands structural editing because the problems are structural — an argument that does not travel, evidence that is not there, claims nobody has checked. The complaint from editors doing this work is consistent: they are paid for the layer the text appears to need rather than the layer it actually needs, and the appearance is precisely what generated text is best at.
Translation went through this earlier and the outcome is instructive. Machine translation post-editing became a standard product line, priced at a discount to translation on the reasoning that the machine had done most of the work. Practitioners have argued ever since that repairing a plausible wrong translation is often slower than translating the sentence, because you must first reconstruct what the source said and then detect where the output quietly departed from it. The rate structure was set by how the work looks rather than by how long it takes, and it stuck. Editing is being offered the same bargain.
Every profession that trains by apprenticeship is fragile in the same place. Copy-editing tests, junior proofreading and reader’s reports were never valuable to publishers mainly as output — they were how somebody learned to see a manuscript. Automating the output while still requiring senior judgement is a plan that works for about a decade and then does not, and no single firm is responsible for the gap because no single firm caused it. Whether the industry builds a replacement route is genuinely undecided, and it is the part of this worth watching rather than predicting.
Staff editors are attached to salaries and notice periods; freelancers are attached to per-project rates that can be renegotiated the moment a client believes the work got easier. That belief moves faster than any evidence about it, so rates fall before anyone has established whether the work actually takes less time. The freelancers best placed to resist are the ones who can point at something a client cannot get from a model — subject expertise, a specific readership, a record of catching the error that would have been expensive.
The shape emerging is not editors doing less but editors doing a different mixture, in which supplying what the model lacks becomes most of the work. Three things sit outside the document and inside the editor: whether the claims are true, what the piece is for, and what it costs if it is wrong. Almost everything defensible about the role is downstream of one of those.
Time spent marking repeated words, inconsistent terminology, undefined jargon and paragraphs that do not follow is time spent on the layer that is genuinely automated now. Run that pass mechanically and arrive at the manuscript with it already done. An editor still hand-marking those and calling it the job is competing where they are weakest, on somebody else’s clock.
The claims, the figures, the attributions and the quotations are where the risk sits, and that risk rose rather than fell once drafting became cheap. It is also the part of the work whose value is easiest to demonstrate afterwards, because a caught error has a name and a consequence, whereas a smoothed paragraph leaves no evidence that anything was prevented.
A model will improve a paragraph without ever asking whether the paragraph should exist, because nobody has told it what the piece is trying to do to whom. Establishing that is a conversation with a writer and a commissioner, it happens before the line work, and it determines which of the model’s suggestions are worth taking at all. It is also the part of editing clients most reliably fail to ask for and most reliably need.
Asked whether a piece is any good, a model produces a balanced assessment, because balance is what it has been shaped to produce. Somebody still has to say that the second half should be cut, that the framing is wrong, or that the piece should not run. Those are the judgements a writer remembers and pays for, and they are unavailable from a system with no stake in the outcome and no reputation riding on being right.
If the durable part of the role is accountability, it is worth stating precisely what you checked and what you did not. An editor who says plainly that they verified the figures, or that they did not and somebody else must, is offering something a model cannot offer and a client can actually rely on. Vagueness about scope was survivable when nobody else was doing any of the work. It is not now.
Concretely: a submission to Kind Channel receives automated editorial notes on its description, and an idea the community votes for receives a feasibility assessment afterwards. The notes block nothing and decide nothing. Community votes determine what gets made. There is a person involved at the feasibility stage, because whether something can practically be produced is a decision somebody has to own — which is the accountability argument above, applied to ourselves.
Nobody was displaced by any of that, and it would be a flattering story to imply otherwise. Kind Channel has never employed a staff of editors reading submissions, so the automated notes did not take a person’s job. They replaced nothing, which is the honest description. A submitter who would previously have received no response at all now receives structured feedback of a limited kind. That is the real comparison, and it is a much smaller claim than replacing an editor.
The limits are worth naming. Kind Channel is new, the community is small and nothing has aired. The automated notes assess a short description against stated criteria: they do not verify that anything in it is true, they do not know your intended audience, and they have no idea what the idea would cost to make. Those are the three gaps described throughout this page, present in our own tooling. If your idea is strong and the description undersells it, the review will not rescue it — and if the description is polished while the idea underneath is thin, the review will not catch that either. A reader who votes might; that is the part of the process doing the work no model here is doing.
Some are, unevenly. Proofreading and the mechanical half of copy-editing were already partly automated before language models and are now largely so, with the losses concentrated in entry-level and freelance versions of that work. Structural editing, fact-checking and commissioning are not going the same way, because they depend on knowing whether claims are true, what a piece is for, and who is answerable for publishing it. The uncomfortable part is that the exposed jobs are the ones the profession used to train people in.
Yes, with a different emphasis than a decade ago. The defensible part of the work is judgement about things outside the document — subject knowledge deep enough to catch a wrong claim, a real feel for a particular readership, and the willingness to say a piece should not run. Training that consists mainly of style-guide mechanics is training for the most exposed layer. The harder practical problem is that the junior roles which used to supply that training are the ones being removed, so the route in has to be assembled more deliberately than it once did.
For low-stakes text making no factual claims, often yes. For anything where a mistake carries a cost, no — and not because the sentence-level work is poor. A model cannot tell you whether the figure in paragraph four is current, whether the person quoted actually said that, or whether a particular phrasing will produce a complaint from someone named in the piece. It also cannot be held responsible for the result, and publishing carries a responsibility that has to rest with somebody who can be wrong at their own expense.
Less often than the pricing assumes. Repairing a fluent draft that is wrong in unmarked places can take longer than writing it, because the editor gets no surface signals about where to look and has to verify everything at the same rate. Machine translation went through this first: post-editing was priced at a discount to translation, and practitioners have argued ever since that the discount reflected how the work looks rather than how long it takes. Whether it is genuinely cheaper depends almost entirely on how much of the draft has to be true.