Why the job is splitting, not vanishing
The honest answer to “will AI replace writers” is that the work is being pulled apart task by task. Language models are good at the part people picture when they think of writing: producing clean sentences on a known subject at speed. They are far weaker at the parts that decide whether a piece is worth publishing at all.
Look at the task list above. Writing original material for publication is one task. Choosing the subject, finding the angle and researching it through interviews, documents and travel are separate tasks, and they sit upstream of the keyboard. A model can produce a draft about a town hall meeting. It cannot sit in the room, notice who went quiet, and decide that is the story.
Revising drafts is the other hinge. Writers rework copy after notes from an editor, a client or a legal reader. Machines produce revisions quickly, but somebody still has to judge which note to take, which to argue with, and what the piece is actually for. That judgment is what buyers pay for, and it is why the pressure shows up first in rates and in junior openings rather than in whole roles disappearing. The Bureau of Labor Statistics projects employment for writers and authors to change by about -0.3% between 2025 and 2035, from roughly 47,940 jobs, with median pay of $76,910 (BLS, 2025). That is a flat market, not a collapsing one.
Where the tools take over and where they only assist
On the measured task time, tools can run roughly 21% with little supervision. That share covers the mechanical end: turning notes and an outline into continuous prose, drafting short promotional and explanatory copy, reformatting the same material for a different length or channel, and organizing research into a usable structure.
About 79% of the task time is shared work, where a person leads and the software speeds up a step. Researching a subject is a good example: a model can summarize filings and transcripts, but the writer decides what to chase and who to call. Working with editors, clients, illustrators and publishers is similar. Drafts move faster; the negotiation over what the piece says does not.
No task on this job’s list is marked as people-only. That sounds alarming until you read what it means: nearly every writing task now has some machine-assisted step, so the human contribution is folded into shared tasks rather than sitting in a protected column of its own. You can see how we split those groups on the Can AI do it? page.
What has actually been tested
Here the page has to be blunt. There is no direct, published test of AI output against the work of professional writers and authors on the tasks this job is actually paid for. Our evidence grade for quality parity is D, which means not measured, so we publish no parity number for this occupation at all.
Benchmarks that score fluency or grammar do not settle it. A real test would need commissioning editors to rate finished pieces blind, with the same brief, the same deadline and the same fact-checking standard, across several formats: a reported feature, a long nonfiction chapter, a client brand piece. Acceptance rates and revision counts would tell you more than any style score. Until something like that exists, treat confident claims in either direction as opinion. Our rules for grading evidence are on the Is it better than a person? page.
Good to know: a missing grade is not a quiet pass mark; it means nobody has run the comparison properly yet.
When the balance could shift
Most likely between 2036 and 2044 (8 in 10 of our scenarios). Two things could pull that window earlier. The first is buyer tolerance: if clients and publishers accept machine drafts with light human editing for mid-tier work, the paid task mix shrinks before the technology improves. The second is cost. Running a model for a year costs a fraction of a writer’s salary, and no robotics are needed here, because none of the work is physical. There is no hardware bottleneck to wait for.
Two things hold it back. Accountability is one: named bylines, legal exposure and brand risk mean somebody has to stand behind the claims, and models still produce confident errors. Access to the raw material is the other. Interviews, embargoed documents, lived experience and trust built with sources are not in a training set. You can read how we build the dated range on the When could it be replaced? page, and the wider method sits at how we score jobs.
How to stay needed
Lean into the tasks that carry risk and originality. Reporting and primary research, where you gather material nobody else has. Developing the idea and the structure, deciding what the piece argues before a word is drafted. And the client-facing side: taking a vague brief, pushing back on it, and defending the finished draft to an editor or a legal reader.
Two skills matter most. First, editing at a professional standard, including fact-checking and sourcing, because reviewing machine drafts is becoming a paid task in its own right. Second, subject depth. A writer who knows construction law, oncology or municipal finance is hired for the knowledge, with the prose as the delivery method.
Writers often look sideways when the market tightens. Three close jobs sit nearby: Editors, Technical Writers and Poets, Lyricists, and Creative Writers. Each has a different task split, so put two side by side on our compare tool before you commit to a move.
Media and communication workers as a group face the same squeeze on routine output, and the information sector page shows how it lands across publishing and broadcasting employers. If you want the blunt version, see what the chatbots themselves say about this job on what the AIs say.