Anthropic has begun implementing statistical text watermarks on its Claude models as part of its compliance efforts for the European Union AI Act. However, a technical detail revealed that on older models like Claude Opus 4.6 and Claude Sonnet 4.6, the watermark is disabled if the 'temperature' sampling parameter is adjusted from its default. Newer architectures, such as Claude Opus 5.5, apply watermarking regardless of sampling configurations. This behavior stems from how the watermarking mechanism, similar to Google DeepMind's SynthID, interacts with sampling parameters that alter the token generation probabilities. AI
IMPACT This technicality could impact compliance reporting for users of older Claude models who customize sampling parameters.
RANK_REASON The article discusses a technical detail about a feature (watermarking) of an existing AI model, rather than a new release or significant industry event.
Read on dev.to — Anthropic tag →
- Anthropic
- Claude Opus 4.6
- Claude Opus 5.5
- Claude Sonnet 4.6
- European Union AI Act
- Google DeepMind
- SynthID
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →