Can Turnitin or GPTZero Detect Watermarked AI Text?

Not as of August 2026. No major AI detector has announced the ability to read Anthropic’s Claude watermark.

Detectors and watermarks work through completely different mechanisms, so the ability to do one does not give you the other.

Why detectors cannot simply read a watermark

They were built to do a different job. AI detectors analyse statistical properties of text, such as word choice, sentence rhythm, and predictability, then output a probability.

Watermark checking is a lookup. It requires knowing the specific signal a provider embedded, and having a tool built to read it. A detector has no access to Anthropic’s scheme unless Anthropic provides a way to check.

What each approach can tell you

• A detector guesses whether writing looks AI-generated. It can be wrong about human writing and wrong about AI writing.

• A watermark records that text passed through a specific model. It carries no probability.

Turnitin’s own documentation says its model may misidentify human, AI-generated, and AI-paraphrased text, and that its result should not be the sole basis for adverse action. That limitation is inherent to pattern-based detection, and a watermark does not fix it. See how Turnitin’s AI detection works.

Why watermarks will not replace detectors

• Coverage. Only participating providers add a mark. Text from an unmarked model carries nothing.

• Absence proves nothing. No watermark could mean human-written, an unmarked model, or a degraded mark.

• Durability is unknown. Reporting says a watermark may persist through some editing. How much editing removes it has not been documented.

So even if watermark checking becomes widely available, pattern-based detection will still be doing most of the work, with all of its existing accuracy problems.

What this means if your work gets flagged

Nothing here changes the practical advice.

• An AI flag is a signal, not a verdict. Turnitin says so itself.

• Your institution’s policy is what actually governs your situation.

• Drafts and version history remain the most useful evidence of your process.

• Ask what specifically was flagged, and on what basis.

Our guides on AI detection false positives and what to do if your essay is flagged go deeper.

Where Lunchbreak helps

Since detectors are what actually get run on your work today, the useful thing is knowing what they say before someone else does.

The AI checker runs your writing against major detectors at once and shows each result separately. That matters because detectors disagree regularly: one flags, another does not.

The humanizer rewrites for tone and flow when your prose reads stiffly. You review the rewrite against your original and keep what you want.

To be direct about the limits: Lunchbreak does not read watermarks, and no public tool does yet. No rewriting tool can promise a particular detector result either, since detection is probabilistic by design.

Check my writing free