When AI writes our drafts, we rarely edit them, and often prefer its wording to our own.

This explores how much people actually revise AI-written drafts before they reach readers. The short answer is: very little. The corpus has one direct measurement, and it is striking. Writers edited AI-generated paragraphs only 23% of the time, and even those edits left the text about 96% similar to what the AI produced Do writers actually edit AI-generated text before publishing?. In practice, the AI's version is usually the published version. The same study tracked a side effect: whatever slant or opinion the AI added goes straight through to the audience.

The more surprising part is why people don't edit. They mostly aren't being lazy. They tend to like the AI version better. In 4,503 cases, writers picked the AI rewrite over their own original paragraph 63% of the time, and about half said the AI version expressed their views *better*, even though the AI versions consistently shifted the original stance Do writers actually prefer AI-edited versions of their own text?. Low editing looks less like neglect and more like approval. Ownership seems to work the same way. People feel the text is theirs when they had a hand in steering it, and a model tuned to their personal style doesn't add to that feeling Does user control over AI text shape feelings of ownership?. A light touch may be enough to make someone feel like the author.

The next stage, readers, doesn't catch what the writers let through. People spot AI content at roughly coin-flip accuracy across text, images, and voice Can people reliably spot content made by AI?. That holds even when AI text measurably differs from human writing in word variety, and trained linguists miss it too Can humans detect AI text if machines can measure it?. So neither writers nor readers reliably filter it. The scale makes this matter. An Internet Archive analysis estimates that about 35% of new websites were AI-generated or AI-assisted by mid-2025, alongside a drop in how varied their meaning is and a shift toward more positive tone How much of the internet is AI-generated now?.

One deeper angle explains why barely-edited text might still be a problem even when it reads fine. AI writes for the person typing the prompt, not for an imagined public. When that text is published without revision, it reaches readers it was never written for Does AI writing collapse the author-to-public relationship?. Editing is traditionally where a writer turns private drafting into public address, so skipping it also skips that step. Readers seem to sense this. They rate disclosure of AI use as more necessary than writers do, especially when AI text is pasted in directly Do readers and writers differ on AI disclosure necessity?.

A caveat: the collection has essentially one study that directly measures edit rates. It doesn't yet show how editing varies by field (journalism vs. marketing vs. academic writing), by writer expertise, or over time as people get used to these tools. The 23% figure is a well-supported starting point, not a settled average.