Anthropic details how Claude's EU-mandated watermarking will work — and where it won't

Anthropic on Friday published a detailed blog post explaining how it will watermark text generated by its Claude chatbot, a move announced earlier this week to comply with the EU AI Act's Transparency Code. The post addresses the practical questions developers and users have been debating since the announcement: How does the watermarking actually work? Can editing defeat it? And what about code?

The company confirmed it will use SynthID-Text, the watermarking approach Google DeepMind detailed in 2024, and said it plans to release a detection API. The system works by introducing imperceptible statistical patterns into low-stakes word choices — like picking "overcast" instead of "grey" — that only someone with the decoding key can detect.

Anthropic was quick to distinguish watermarking from the kind of statistical "tells" that third-party AI detectors search for, such as characteristic phrasing patterns. "Picking up on these patterns is fundamentally different from checking for a watermark," the company wrote.

Will editing remove the watermark?

Anthropic acknowledged that a determined rewrite can defeat the watermark, but said "light editing probably won't remove the watermark completely." Only a complete rewrite where every word is replaced will reliably strip it out — at which point, the company argues, it's debatable whether the text is still AI-generated.

For text that Claude only proofreads or lightly edits, the watermark may be nearly absent, since "nearly all the words" would still be authored by the human, leaving little for the watermark to attach to.

Code gets a lighter touch

One question developers will care about: Does watermarking affect generated code? Anthropic says code will carry less of a watermark than prose, because the model must produce working code and has less freedom to choose between equally valid options. However, the watermark can be applied in places with arbitrary choices, such as comments within code. "By definition," the company said, "it will have a negligible effect on the actual code produced."

The announcement has already sparked backlash on Reddit and X, with some users threatening to cancel subscriptions. But Anthropic stressed that Claude won't be alone — other major model developers have signed the same Code of Practice and will be implementing their own watermarks.