Claude Watermark: Anthropic Explains Tech as Users Cancel

Anthropic has published new details on its text watermark for Claude. The system uses a SynthID-style approach: when several phrasings are equally good, the model makes subtly steered word choices that together form a statistical pattern — invisible to readers, detectable only with the right key.

Per TechCrunch, Anthropic stresses that output quality does not suffer: to a reader, a watermarked response is indistinguishable from an unwatermarked one. Light editing does not fully remove the pattern; only rewriting essentially every sentence erases it. Code carries minimal watermarking, since word choice there is functionally constrained.

A detection API is planned, letting third parties check whether a text came from Claude. Anthropic points to transparency obligations such as the EU AI Act's code of practice, which requires AI-generated content to be identifiable.

Part of the user base is pushing back regardless: Business Insider reports dozens of subscribers canceling Claude Max, worried that watermarked text could expose them as AI users at work or school. On Reddit, some accuse the company of acting covertly.

The clash exposes a core tension: what regulators and platforms celebrate as overdue transparency, paying users experience as losing control over their own text.