Has this text been
watermarked?

The short answer

It depends which of four very different mechanisms you mean. People use “AI watermark” for statistical word-choice watermarks, hidden Unicode characters, signed provenance metadata, and post-hoc AI detectors — and each lives in a different layer of the text, is checked a different way, and tells you something different. Two of the four you can check yourself today; the other two you cannot.

1. Statistical watermarks — you cannot check these yourself

Systems like Google’s SynthID-Text (deployed in Gemini) and Claude’s announced watermark carry their signal in the pattern of word choices made during generation. Nothing is added to the text, so no amount of reading, character inspection, or file analysis will reveal one. Detection requires the provider’s secret key: the detector re-scores every word choice against keyed pseudorandom functions and asks whether the correlation is stronger than ordinary text would produce.

Practical consequence: only the provider (or whoever they license the detector to) can test for these. Anthropic has announced a detection API but described it as forthcoming. Any third-party service claiming to detect Claude’s watermark today is overstating what is possible. Read what Anthropic actually announced for the specifics.

2. Hidden Unicode characters — check these in seconds

Older marking tricks, and plenty of accidental copy-paste artefacts, live in invisible code points: zero-width spaces, word joiners, soft hyphens, directional marks. These are the opposite of a statistical watermark — trivially visible to software, invisible to readers.

To check: paste the text into an editor with invisible-character rendering switched on, search for code points like U+200B, or run the draft through our Character cleanup, which reports exactly how many characters it adjusted. Zero adjustments means the text contained none. Our invisible-characters guide lists the specific code points and safe removal practice.

3. C2PA provenance credentials — file-level, fragile for text

C2PA content credentials are cryptographically signed metadata attached to a file, recording how it was made. They are verified with a C2PA-aware tool, not by inspecting the prose. For text they are fragile in an instructive way: the moment someone copies the words out of the original file and pastes them anywhere else, the credential is left behind. Absence of a credential therefore proves nothing about origin.

4. Post-hoc AI detectors — usable by anyone, reliable for no one

Detectors like GPTZero or Turnitin’s AI score need no key and no cooperation from the model provider — they estimate AI likelihood from learned linguistic features. That accessibility is their weakness: they are wrong in both directions often enough that OpenAI withdrew its own classifier in 2023 citing low accuracy, and a Stanford study found several detectors disproportionately flagged text by non-native English writers as AI-generated.

Treat a detector score as one weak signal among several, never as proof — in either direction.

A practical checking order

• Run a character inspection first. It is the only check that gives a definitive answer, and it takes seconds.
• Check file provenance if you have the original file. A valid C2PA credential is strong evidence about that file’s history.
• Use provider detection where it exists. For statistical watermarks, only the provider’s own detector is meaningful.
• Weigh AI detector scores last, and lightly. Corroborate with context — version history, drafts, the author’s other writing — before drawing conclusions.