Can you check text for an AI watermark?

A text watermark is a slight tilt in word choices, set by a secret key. To find it, a detector scores every word of a text with that key and checks whether the favoured choices turn up more often than chance would allow. Without the key there's nothing to score: watermarked and unwatermarked text read the same, and a single text gives an outsider nothing reliable to measure.

Even with the key, the answer is a likelihood, not proof. Detection gets weaker on short texts and after the wording has been changed.

• Claude · Anthropic

  A detection API in private preview, for organisations such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups.

• ChatGPT and Codex · OpenAI

  A detector for approved researchers and expert organisations, by application.

• Gemini · Google

  The SynthID Detector portal checks text, images, audio and video made with Google's AI, for journalists, media professionals and researchers on a waitlist. The check in the Gemini app covers images, video and audio, not text.

If you work for one of those organisations, apply to the provider. Otherwise there's no way to run the check yourself.

Search for an AI watermark detector and you'll find tools that promise to check Claude or ChatGPT text. None of them holds a provider's key, so what they test is one of two other things:

• Hidden characters, such as zero-width spaces and unusual spaces. Worth knowing about, but they aren't the watermark.
• Writing style, which is what AI detectors judge. A guess from style can't see a watermark; it can flag human writing by mistake and miss AI text that has been edited.

Neither tells you whether a provider's detector would find its watermark.

Hidden characters are the part you can see for yourself. Paste a text into glyphwash and the result lists every hidden character it found: what it was, where it sat and what replaced it, including whole messages spelled out in invisible tag characters. The invisible characters AI text can carry covers each kind.

For the watermark itself, the practical question is where the text came from. Assume it's there in text from a Claude model released after 2 August 2026, from Gemini in the Gemini app or on the web, and from ChatGPT in the EU as OpenAI's rollout reaches it. The Claude watermark and the ChatGPT watermark go into the details.