Anthropic pledges to embed watermarks to help discern AI slop in sop to EU

Anthropic, an artificial intelligence company, has announced plans to embed watermarks in the text and files generated by its upcoming models in the European Union. This move aims to comply with content and transparency rules outlined in the AI Act. The company revealed these plans in a help document published on Monday.

In addition to the newly launched models, Anthropic is also working on embedding watermarks into the output of its existing models during a transition period permitted by EU law. This watermarking process will apply to generated content from supported models across all platforms where Claude is available globally.

Anthropic's decision to watermark their AI-generated content may boost the appeal of open weight models, potentially alienating some current Claude users who may not want consumers to be aware of the AI's provenance. Claude users have previously expressed dissatisfaction with pricing changes, reliability issues, and model safeguards, which have impacted legitimate work for safety reasons.

Claude users have raised doubts about the effectiveness of text-based watermarking, as researchers have shown that image-based watermarking can be removed. Anthropic plans to apply watermarks to output from covered Claude models on various platforms, including API, Claude, Claude Code, Claude Cowork, and Claude Tag. The company also intends to provide further instructions on detecting these watermarks in future documentation.

When a supported Claude model generates text, it seamlessly incorporates an imperceptible watermark directly into the text itself. This watermark is invisible, does not alter the meaning, quality, or readability of Claude's response, and travels with the text when copied and pasted elsewhere. It may persist through certain editing processes.

Watermarking will be implemented at the model level, ensuring that it remains present regardless of the Claude product or surface from which the text originates. While the exact method of making these watermarks resistant to removal is unclear, it is expected that an optical character recognition system could potentially strip or omit these obscure marks from Claude-generated text.

Despite Anthropic's assurance that their text marking scheme does not alter the meaning of Claude's output, the company acknowledges that detected marks may not conclusively prove Claude's involvement in content creation. Nonetheless, the scheme may be sufficient to meet legal compliance requirements.