Can AI detectors catch text that's been edited by hand?
Light edits rarely hide AI writing. Heavy rewrites change the picture. What happens to a score as a person reworks an AI draft, sentence by sentence.
Tomás Lindqvist
Machine learning engineer · Sep 3, 2026 · 6 min read
One of the questions we hear most: if someone takes a chatbot draft and edits it, will a detector still catch it? The honest answer is that it depends on how much editing, and the sentence-level view makes that easy to see.
What light editing does
Swapping a few words, fixing the opening line or changing the tone of a paragraph leaves most of the underlying structure intact. The sentences still follow the same rhythms and word patterns, so they're usually still highlighted.
What heavy editing does
When a person rewrites the argument, adds their own examples and restructures paragraphs, the text genuinely becomes mixed authorship. A good detector should report exactly that: some sentences flagged, others clear, and an overall score somewhere in the middle.
- Mostly unchanged AI draft: high AI probability across most sentences.
- Light edit: a few sentences clear, most still flagged.
- Substantial rewrite: a genuine mix, shown sentence by sentence.
- Full rewrite from an AI outline: usually reads as human, because it mostly is.
A mixed result isn't the detector being unsure. Often it's the most accurate description of how the text was made.
What about paraphrasing tools?
Automated rewriters change words but rarely change structure, and they leave patterns of their own. Our latest model was trained on paraphrased text specifically, so these drafts tend to stay flagged. They also tend to read worse, which is often the first thing a teacher or editor notices.
Written by
Tomás Lindqvist, Machine learning engineer
Tomás works on model training and evaluation, with a focus on keeping false positives rare.