Most people barely edit what an AI writes for them, and when they do, they often can't see what's off.

This explores what writers actually do to check and fix AI-drafted text before publishing, and the corpus's blunt answer is: mostly nothing. In one study, writers edited AI-generated paragraphs only 23% of the time, and the edits they did make left the text 96% similar to the original Do writers actually edit AI-generated text before publishing?. Revision is the exception, so the AI's voice reaches readers close to unchanged.

Part of the reason is that writers don't see anything to fix. In a study of 4,503 cases, 63% of writers preferred the AI's version of their own paragraph, and 52% said it reflected their views better, even though the AI versions systematically shifted the stance Do writers actually prefer AI-edited versions of their own text?. The shifts weren't random. Across 29 measured dimensions, AI assistance made writers read as more extreme, confident, agreeable and privileged to their readers Does AI writing assistance change how readers perceive the writer?. Detection doesn't help either. LLM text differs measurably from human text, yet even trained linguists can't reliably spot it, and newer models are harder still Can humans detect AI text if machines can measure it?. A writer rereading a draft for anything that feels off has little to catch on. One reading of the preference result is that writers do the interpretive work of making the text fit them. AI output carries the markers of an utterance without the event behind it, and the human supplies the missing orientation Does AI generate genuine utterances or just text patterns?.

Even a diligent line-edit may not reach the deeper problems. AI fiction can be separated from human fiction with 93.2% accuracy using only story-level choices like character agency and chronological structure. Those choices resist humanization because fixing them takes a rewrite, not a surface polish Can AI stories be detected without analyzing writing style?. Other notes describe the same gap as missing structure: artificial text lacks dialogue with a real interlocutor, continuity with context, embodied authorship and political situatedness Does AI-generated text lose core properties of human writing?. It is also written for the person who typed the prompt, not for the public who will read it Does AI writing collapse the author-to-public relationship?. So the revision question that matters may be who the text is actually for, and that is a different question from whether the sentences read well.

The verification approaches that do show up in the corpus don't rely on rereading. In Data2Story, an Inspector binds every number, quote and asset to its origin, so a reviewer can audit where each claim came from. Human raters favored it, and the notes frame provenance rather than fluency as what lets professional newsrooms adopt agent output Can source traceability make AI writing trustworthy?. In shared editors, writers wanted to see when, where and how their collaborators used AI, partly to verify the AI-written passages, though some found full visibility intrusive Do writers want to see each other's AI prompts in shared editors?.

Two shortcuts look like verification but aren't. Asking another LLM to check the draft is risky, because LLM judges score higher when a response includes fake references or rich formatting, regardless of content quality Can LLM judges be tricked without accessing their internals?. And a citation that looks right proves nothing: one demonstration produced 288 complete finance papers with invented justifications and fabricated citations Can AI generate hundreds of fake academic papers automatically?. The corpus has little on what good personal revision habits look like. What it shows is that unaided rereading fails, and the working checks it describes are traceable sources and visible AI use.