Edition
Anthropic's watermark rollout leaves a loophole on two older Claude models
Anthropic is adding text watermarks to four Claude models, though API users can evade them on two by changing a single sampling parameter.
The temperature, top_p, and top_k parameters control randomness in word choice; Anthropic says a minor change to any one can bypass the watermark and advises keeping the defaults when relying on it for compliance. Anthropic says the watermark adds no characters or tokens, does not affect pricing or request and response formats, and only slightly slows output.
This summary is based on the available full source text. Summary prepared with AI assistance. The original source is authoritative.
Original source
The Decoder
Date stated with the source: 31. 10. 2026
Open the original at the publisher ↗