Editor and proofreader — tasks, one by one

Proofreading and correction

Language models find errors well but over-correct. Evaluations found ChatGPT produces fewer under-corrections but more over-corrections than dedicated tools, and that it cannot effectively correct agreement, coreference and tense errors across sentences. Elsevier says its own teams use AI to support proof preparation and copy editing.

2023-era benchmark studies and one publisher's description of its own process; neither measures proofreading jobs.

Copyediting for accuracy and style

AI edits more but not better. In a comparison on two papers, a university-hosted GPT made about three times as many corrections as a human editor, but only 61% of them were judged improvements; the human took 3.75 and 4 hours per paper. In a study of 1,057 AI-written paragraphs edited by professional writers, experts largely preferred text edited by other experts. US projections expect editor employment to decline 1 percent from 2025 to 2035.

Small studies of academic and literary text; they do not show how publishers staff copyediting.

Structural and developmental editing

The evidence here shows models rewriting heavily rather than editing with judgement: studies find language models favour more generative rewrites that may improve fluency but risk altering nuance, and experts prefer human editors' edits. No record here shows AI doing developmental editing in practice.

Inferred from studies of short texts; no record measures developmental editing work.

Standards and publication decisions

Publishers keep the decision with editors. Elsevier says AI tools cannot replace an editor's critical thinking or final decision-making and that editors remain fully responsible; the Guardian says significant AI-generated elements require the explicit permission of a senior editor.

Two organisations' own policies; others differ, and policies do not measure staffing.