■ HIGH RISK ■ Media & Communication
Substantially, yes. Line editing and copy editing are exactly what language models do best, and the publishing economics were already fragile — what survives is commissioning judgement and accountability.
“AI proofreads and restructures better than most. But editorial judgment and taste? Still a human call.”
Our AI replacement risk score — how we score jobs
Editing spans a ladder of tasks. At the bottom, proofreading and copy editing: grammar, house style, consistency, factual formatting. In the middle, line and structural editing — tightening prose, resequencing arguments, cutting a piece to length without gutting it. At the top, commissioning and developmental work: deciding what is worth publishing, shaping a writer's raw idea into something readable, judging what an audience will actually care about, and taking responsibility when something goes wrong.
The bottom two rungs are the most directly exposed tasks of any knowledge job. Models apply style guides consistently, catch inconsistencies a tired human misses at 6pm, restructure a rambling draft, and cut 1,200 words to 800 without breaking the argument. Publishers and newsrooms have been thinning subediting desks for two decades under cost pressure, and AI is the accelerant on an already-burning building. Freelance copy editing rates have been under visible pressure, and volume content operations have moved to a model of AI drafting with light human review — which is a different, lower-paid job than editing.
The top of the ladder holds better, for two reasons. Taste is a judgement about a specific audience at a specific moment that models approximate blandly, and accountability is legal: someone must decide a story stands up, that a source is real, and that the publication will defend it. Fact-checking has become more important precisely because generated text is confidently wrong. Our risk score of 70 reflects a profession where most of the hours are automatable, most of the jobs are in those hours, and the remaining roles are senior, fewer, and harder to reach.
Automatability: our editorial assessment of current and near-term AI capability
Serious pressure already, intensifying through 2030. Copy and subediting desks have been shrinking for years and AI has accelerated that sharply; freelance copy editing is being repriced downward now. By the early 2030s expect most routine editing in commercial content, publishing, and marketing to be machine-first with human review. Commissioning editors, standards editors, and senior developmental roles survive, but there are far fewer of them and the path up has lost its lower rungs.
For copy editing and line editing, it is genuinely competitive and tireless in a way humans are not at the end of a shift. Where it fails is judgement: knowing this joke will land wrong with this audience, that this paragraph is technically fine but says nothing, or that a claim needs a second source. It edits sentences well and edits arguments poorly.
Only with a clear plan to reach the senior end. Entry-level copy and subediting roles are disappearing, and freelance rates for routine work are falling. Editing remains viable as commissioning, developmental, or standards work, or in specialist fields where accuracy carries legal or financial consequence. Treat generic proofreading as a service, not a career.
Stop competing with software on price for tasks it does well. Reposition toward developmental editing, subject-matter specialism, and verification of AI-generated material — including building the workflow a client needs. Package work as outcomes rather than hourly proofreading. Clients still pay properly for someone who improves the substance rather than the surface.
Commissioning editors, standards and ethics editors, developmental editors working closely with authors, and technical or scientific editors in regulated fields. The common thread is decision-making authority and accountability. The least safe are copy editors, subeditors on routine content, and anyone whose value proposition is consistency and error-catching alone.