Do writers actually edit AI-generated text before publishing?

A common reassurance about AI writing assistance is that humans remain in the loop — they will edit, correct, override. The persona-distortion study tested this assumption directly. Writers were given AI-generated paragraphs and asked to edit them until the text reflected their opinions to their satisfaction. The result: writers edited the AI-generated paragraphs only 23 percent of the time, and most edits were minor — median Levenshtein ratio of 0.96, meaning the edited text was 96 percent identical to the AI's original.

This finding has two implications. First, the standard "human-in-the-loop" defense against AI text quality concerns is empirically wrong at population scale. Editing is rare and shallow when it does occur. The AI's text is reaching its audience in nearly the form the model produced it. Second, this means the persona distortions documented in the same study — opinionated, confident, demographically privileged, emotionally compressed — propagate with minimal human modulation. The distortion is not filtered by the writer's revision; it is embraced or ignored.

This forecloses one common mitigation strategy: relying on the writer to detect and remove distortions before publication. The writer who would have caught and corrected the distortion is the same writer who, the study shows, mostly does not edit and mostly prefers the AI version even after being given the chance to edit it. The distortion arrives at the audience because the writer does not interrupt it. Any intervention that hopes to reduce AI's influence on public discourse cannot rely on the writer-as-gatekeeper assumption — that role, in practice, is not being performed.