An AI detector flagged your writing. Here is what to do
By Claude Watermark Research · Updated
A detector score is a probability estimate about writing style, not evidence of what happened. The strongest rebuttal is process evidence: document version history showing the text being written over time. That is a record of events, and it beats a classifier score in any fair hearing.
Why you were flagged
These tools score how predictable your word choices are against a language model. Plain, uniform, well-structured prose reads as machine-like, which is why the documented false positives cluster on non-native English speakers, on people trained in technical or academic writing, and on anyone who writes simply on purpose.
Turnitin has publicly acknowledged false positives, and a number of universities have restricted or switched the feature off entirely. You are not the first person this has happened to.
What actually proves authorship
**Version history.** Google Docs File → Version history, Word's version pane, or a git log shows the document growing over time, with the false starts and rewrites a finished file cannot show. Export or screenshot it before anything else.
**Drafts and notes.** Outlines, research notes, messages where you discussed the argument, anything timestamped before submission.
**A conversation.** Offer to talk through your argument. Someone who wrote a piece can explain why they cut a paragraph; that is hard to fake and easy to demonstrate.
How to respond
Ask what the score actually was and what threshold the institution treats as meaningful. Many policies require corroborating evidence rather than a score alone, and asking makes that explicit.
Do not rewrite the piece to lower a score before the meeting. It reads as consciousness of guilt and it destroys the version history that was your best evidence.
If artifacts are part of the accusation — a `claude` class name in the markup, invisible characters — that is a separate and much more concrete question than a style score. It shows text passed through a chat interface, but it cannot distinguish writing from proofreading.
Questions
- Can I prove a negative?
- Not directly, and you should not have to. The burden sits with the accusation. Process evidence is what shifts it.
- Does running my text through a checker help?
- It answers a narrower question: whether your document contains artifacts anyone can point at. That is worth knowing before a meeting, whichever way it comes out.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
