Anthropic expands watermark to older Claude models with a loophole
In brief
- Anthropic expands its text watermark to four Claude models, including two older ones (Opus 4.6 and Sonnet 4.6) where it can be bypassed.
- The watermark uses a statistical pattern in word choice without altering quality or adding characters.
- Anthropic cites the EU AI Act as the reason, but questions persist about its effectiveness and legal implications.
Anthropic is rolling out its text watermark to four additional Claude models to meet the EU AI Act's December deadline. However, API users can bypass the watermark on two older models, Opus 4.6 and Sonnet 4.6, by adjusting a single sampling parameter.
The company describes the watermark as an "imperceptible pattern in word choice" that does not affect meaning, quality, or readability. It adds no extra characters or tokens, has no impact on pricing or request formats, and only minimally slows output speed. Anthropic emphasizes that the watermark contains no user or conversation data. Yet, for the two older models, the watermark only activates if default sampling settings (temperature, top_p, top_k) are used.
Anthropic frames the rollout as necessary for compliance with the EU AI Act, noting the watermark is model-level and works across all hosting platforms. While similar methods exist from Google DeepMind and OpenAI, concerns remain about potential text quality degradation and legal complications for law firms using AI tools.
What it changes for you
- LegalLaw firms must verify if their AI tools generate watermarks, risking exposure of AI use despite client bans.
This summary gives you the essentials. The full article, its figures and quotes are on The Decoder.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error