Anthropic to watermark Claude text output under EU AI Act, with a sampling loophole
Overview
Anthropic says future Claude models will embed a watermark in generated text to comply with the EU AI Act, which requires AI providers serving the EU market to mark AI-generated content from August 2.
O'Reilly Radar reports that the mark covers all Claude outputs, including Claude Desktop, Claude Code and the API, and uses a steganography method based on SynthID-Text, an open-source Google DeepMind project. Anthropic says the mark carries no information about the user.
Anthropic later told API customers it will also watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, tying the rollout to the Act's December deadline, according to The Decoder. The Decoder reports that on Claude Opus 4.6 and Claude Sonnet 4.6 the watermark applies only when the temperature, top_p and top_k sampling parameters stay at their defaults, and that changing any one of them removes it. Customers cannot turn the watermark off.
Written by AI from the articles below · updated Oct 31, 4:14 PM ET
Check the sources:
Developments
2 developments
- Oct 31, 3:20 PM ET · 1 articleAnthropic expands text watermark to four more Claude models, with a loophole on two older onesThe Decoder: Anthropic extends text watermark to four Claude models, with a loophole on two
- Oct 22, 6:55 AM ET · 1 articleAnthropic's Claude models will embed AI text watermarks to comply with the EU AI ActO'Reilly Radar: Anthropic's Claude models will embed AI text watermarks to comply with the EU AI Act
Article timeline
The articles in this story. Times are ET.
- The DecoderNewsAnthropic extends text watermark to four Claude models, with a loophole on two
AIAnthropic says it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6 in an email to API customers. On Claude Opus 4.6 and Claude Sonnet 4.6, changing any one of the temperature, top_p, or top_k sampling parameters from its default removes the watermark. Anthropic ties the rollout to the EU AI Act's December deadline.
- O'Reilly RadarBlogAnthropic's Claude models will embed AI text watermarks to comply with the EU AI Act
AIAnthropic says future Claude models will embed a watermark in generated text to comply with the EU AI Act, which requires AI providers serving its market to mark AI-generated content from August 2. The watermark applies to all Claude outputs, including Claude Desktop, Claude Code, and the API, and covers listed models such as Claude Opus 5.5 and Claude Sonnet 5.5. The watermark uses a steganography method based on SynthID-Text, an open-source project from Google DeepMind, and carries no information about the user.
Heat Trend
Not enough continuous observations to show a trend yet.