Detectors vs watermarks

Do GPTZero, Turnitin, or ChatGPT detect Claude watermarks?

Short answer: not the Claude watermark. GPTZero, Turnitin, Originality.ai, and similar products classify AI-like writing. Anthropic’s mark is a sampling bias in Claude word choice. ChatGPT cannot read that mark. Anthropic has not shipped a public consumer detector.

Check inspectable tracesHow Claude watermarks work →

Two different signals

GPTZero / Turnitin-style classifiersClaude text watermark
What it scoresWhether prose looks machine-writtenWhether token choices match a hidden pattern
Who holds the keyThe classifier vendorAnthropic (not published as a public API)
False positivesWell documented on human textDifferent failure mode; short text is weak either way
What this site can inspectNot their modelsInspectable traces only — see the checker

GPTZero

GPTZero and cousins estimate “this sounds like a language model.” Heavily edited Claude drafts can still look AI-ish to a classifier, or look human after a strong rewrite, regardless of whether Anthropic’s watermark is still in the tokens. A green GPTZero dashboard is not proof the statistical mark is gone. A red one is not proof it is Claude.

Turnitin

Turnitin’s AI-writing indicator is a school integrity product, not a Claude-watermark product. It does not mean your university is running Anthropic’s verifier. If you are a student, read Claude watermarks on essays for what this tool actually does — and what it will not promise.

ChatGPT (and other chatbots)

Pasting Claude output into ChatGPT and asking “is this watermarked?” is not a detection method. The other model has no access to Anthropic’s watermark key. It may guess, moralize, or invent a procedure. That guess is not a check.

What you can do instead

  1. Use the in-browser checker for HTML class names, zero-width characters, and exotic spaces.
  2. Understand clean vs rewrite: cleaning is free; disrupting the statistical mark takes paid paraphrase.
  3. Do not treat classifier scores as a verify loop. When Anthropic ships a public detector, a detect → rewrite → re-check loop becomes possible. That loop does not exist yet.

FAQ

Does GPTZero detect Claude watermarks?

GPTZero scores whether writing looks AI-generated. That is a different signal from Anthropic’s statistical watermark in Claude token choice. A GPTZero score is not a Claude watermark detection.

Does Turnitin detect Claude watermarks?

Turnitin’s AI-writing reports estimate machine-like prose. They are not Anthropic’s watermark verifier. Schools using Turnitin are not running a public Claude watermark detector.

Can ChatGPT tell if text is watermarked by Claude?

No. ChatGPT is another model. It cannot read Anthropic’s watermark key, and asking it to “detect a Claude watermark” is not a real check.

Open Claude watermark removerWhat the text watermark is →

Related guides

Not affiliated with Anthropic, GPTZero, Turnitin, or OpenAI. We do not operate a public Claude watermark detector. Updated August 30, 2026.