Will Turnitin detect text written with Claude?
By Claude Watermark Research · Updated
Turnitin's AI detector does not look for Anthropic's watermark and has no access to it. It is a statistical classifier trained to spot text whose word choices are more predictable than typical human writing. That means it can flag text Claude never touched, and miss text Claude wrote entirely — the two systems are unrelated and neither can confirm the other.
What Turnitin is actually measuring
The detector scores how predictable each sentence is against a language model. Human writing tends to be uneven — odd word choices, abrupt shifts, redundancy. Model output is smoother because it is drawn from the high-probability part of the distribution. Turnitin reports the share of a document that reads as machine-smooth.
That is inference from style, not evidence of authorship. Nothing in the process identifies which model produced anything, and nothing detects a watermark. It would return a score on a document written in 1950.
The false-positive problem is documented, not theoretical
Turnitin has publicly acknowledged false positives, and independent testing has repeatedly flagged writing by non-native English speakers and by people who write in a plain, uniform style. Those are exactly the traits the classifier reads as machine-like.
This is why a growing number of universities have restricted or disabled the feature. A score is a probability estimate about style, and it cannot be verified by anyone, including Turnitin.
Why the watermark does not help either side
Anthropic's watermark is detectable only with the secret key used at generation, which Anthropic holds and has not published a detector for. Turnitin cannot check it, no university can check it, and neither can we.
Practically: as of August 2026 no shipped Claude model carries the mark at all, so even if a detector existed there would be nothing to find. The one artifact that is genuinely verifiable is far more boring — the HTML class names the Claude web interface pastes into your document, which anyone can see by inspecting the file.
Questions
- Can Turnitin tell which AI model was used?
- No. It produces a single likelihood score for machine-generated text. It does not identify a model, a version or a provider.
- Does editing AI text lower the Turnitin score?
- Usually, because editing makes word choice less predictable. But since the score is an estimate rather than a measurement, no rewrite can guarantee a particular result.
- If Turnitin flags me and I did not use AI, what proves it?
- Version history is the strongest evidence — Google Docs revision history or a git log shows the document being written over time. A detector score cannot rebut that.
Check your own text. Free, unlimited, no account, and it runs in your browser so nothing is uploaded.
Open the checker
