Claude's watermark and Turnitin: what schools can actually see

Whether plagiarism checkers can read Claude's watermark, what they do instead, and why a flag is weaker evidence than it looks.

Removing the watermark · 2 min read

If you are a student wondering whether your university can read Claude's watermark, the answer today is no — and the reason matters more than the answer.

Turnitin cannot read Claude's watermark

Anthropic has not published its detection mechanism. Without it, no third party has a detector — not Turnitin, not GPTZero, not Copyleaks, not your institution's IT department.

What those tools run instead is statistical AI detection: analysing the text's own properties and guessing. That is a completely different thing from reading an embedded mark, and it is much weaker.

What AI detectors actually measure

Two signals, mostly:

Perplexity — how predictable the text is to a reference language model. Model output tends to be less surprising than human writing.

Burstiness — how much that predictability varies across the document. Human writing swings; model output is flatter.

Neither reads a watermark. Both are inferences about style.

Why a flag is weak evidence

The false positive rate is high, and it is not random about who it hits:

• Non-native English writers. A smaller active vocabulary and more regular grammar genuinely lowers perplexity. This is the best-documented failure mode, and it falls hardest on students least able to contest it.
• Technical and scientific writing, which is supposed to be formulaic.
• Well-edited writing. Copy editing removes exactly the surprising choices that raise perplexity. A polished draft scores worse than a rough one.
• Short submissions, where there is not enough text for any statistical measure.

Run one passage through three detectors and you will often get three incompatible scores. None is measuring ground truth.

And when a detector can read a mark

One case exists: Google's SynthID, in Gemini, where Google holds the key. Even there, the output is a confidence score that degrades under editing, translation and short length.

Anthropic's mark will presumably become checkable when they publish detection documentation. When that happens, note what Anthropic says it would prove: content may have been processed by Claude — not that Claude authored it. Text you wrote and asked Claude to proofread carries the same mark as text Claude wrote from scratch. That is a critical distinction for anyone accused on the strength of one.

If you have been flagged

In rough order of usefulness:

1. Ask which tool and what threshold. A specific number from a named tool is contestable; "the detector said so" is not.
2. Produce process evidence. Version history, drafts, document revision timelines. Far stronger than arguing about the text, and the reason to keep drafts.
3. Ask for the same tool to be run on older work of yours that is definitely human. If that also scores high, you have established how the tool reads your voice.
4. Cite the false positive research on non-native writers. It is published, not anecdotal.

The thing no tool solves

If your institution requires you to disclose AI assistance, that obligation stands regardless of what any watermark or detector does. Removing a mark does not change what you did, and no text tool resolves a disclosure requirement. Worth being clear-eyed about that before deciding what to submit.

Related: why detectors flag human writing · how detectors work · does Anthropic watermark Claude's text?