Anthropic enables watermarking for everyone. Turns out invisible text is free

Anthropic has quietly begun rolling out watermarks in text generated by Claude models. The move is tied to the European Union Code of Practice, but the company is not limiting it to European traffic. It does not want to distinguish request sources, and the marginal cost of watermarking is close to zero.

The signature is created during token selection

The method developed by Scott Aaronson and Hendrik Kirchner uses the randomness already present in generation. The model still assigns probabilities to possible tokens, but selection is guided by a pseudorandom source derived from a secret key. Given enough text, a score can show how closely those choices fit that source without a practical change in output.

Google tested the impact across 20 million cases

Google has been rolling out the same principle since 2024 and also offers a public detector. In a test with a sample size of 20 million, it found no difference in user feedback. That supports the claim that the signature need not reduce text quality or user satisfaction.

Rewriting weakens the trace

Detection needs a sufficiently long sample, and the layer is not indestructible. Rewriting in your own words can remove it, while partially preserved text retains the signature in proportion to how many original choices remain. Anthropic says it will provide an API for verification.

A new standard for future releases

The major Western AI labs signed the Code of Practice, which calls for similar marking in future models. Anthropic is therefore not an isolated experiment but the first visible part of a broader commitment that Google already follows and OpenAI also says it intends to meet.

Lilith's verdict

A watermark cannot manufacture trust by itself. Until Anthropic releases its promised verification API, people will see only a signature that its author knows how to read.