Anthropic says it will watermark text generated by its AI models | TechCrunch

Artificial intelligence safety and research company Anthropic has announced plans to implement digital watermarking across its suite of AI models, including its popular Claude chatbot family.

This strategic update is designed to align the company's products with the European Union's recently enacted AI Act, specifically compliance measures under the Transparency Code that require clear identification of synthetic content. Under the new protocol, all AI models launched by the company after the initiation date in August will automatically embed identifying signatures within both generated text and exported file formats. The company also intends to integrate this tracking mechanism into its legacy models over time, ensuring a uniform standard across its entire portfolio of developer tools, consumer interfaces, and enterprise software.

The watermarking mechanism is engineered at the model level, which means that the identifiers are woven directly into the generated text itself. Consequently, these digital markers are designed to persist even when text is copied and pasted or subjected to mild editing processes. For document and media files, Anthropic is leveraging the widely recognized C2PA open standard to flag machine-generated assets. This initiative reflects a growing trend among leading technology firms to adopt robust identification techniques as governments globalize regulatory frameworks for artificial intelligence. By introducing these features, Anthropic joins a coalition of major industry players committing to enhanced content transparency to mitigate risks associated with synthetic media.