Since the EU AI Act's transparency rules took effect in August 2026, AI providers serving the EU have had to mark generated content. Anthropic now hides a watermark in Claude's word choices, using a version of Google DeepMind's SynthID-Text. Google has used SynthID-Text in Gemini since 2024. Here's what a watermark can and can't tell anyone, based on Anthropic's own documentation.
What a watermark can show
- That a specific model was likely involved. Anthropic says the check answers one question: "What is the likelihood this was partly written by Claude?"
- That copy and paste didn't wash it out. The mark lives in the words themselves, so it travels with them.
- That text was translated by the model, since every word in a translation is the model's choice.
What it can't show
- Who. "Watermarking carries no identifying information and can't be traced to a specific person, organization, or chat."
- How much. It "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'"
- Ownership or authorship. Anthropic says a watermark "doesn't say anything about ownership or authorship."
- That there was no AI. Short passages, heavy rewrites, older models, other vendors' models and text from cloud partners where marking isn't on yet may all show nothing. A missing mark proves nothing.
- Light proofreading. If the model only fixes grammar, there's very little for the mark to attach to.
Watermarks vs. "AI detectors"
Detector tools guess from style "tells." A watermark check uses the provider's private key. They are different tests and shouldn't be treated as the same evidence. Right now, Anthropic's detector is limited to regulators, media, fact-checkers, researchers and companies with their own compliance duties.
The practical rule
Treat a watermark as a signal, not a verdict. If you use AI, and you probably should, know what your tools do, disclose it where it matters, and keep a human who stands behind the final text.