Anthropic unveils AI watermarks to comply with EU law

Anthropic, an American AI company, plans to introduce an invisible watermark to text generated by its Claude chatbot to adhere to EU regulations. However, the company clarified that this mark only signifies the text was processed by Claude, not that the chatbot created it. The new feature, unveiled on Monday in San Francisco, remains undetectable to humans and persists even after text modification.

Yet, Anthropic warned about certain limitations. Detecting a Claude watermark merely indicates the text was processed by Claude, but does not confirm the source. For instance, Claude may not be the original author; people frequently use Claude for tasks like proofreading, translation, summarization, or file conversion. The output may include a Claude mark even if the original ideas, text, or data came from another source.

Additionally, short texts or heavily edited ones might not carry a watermark. Claude will embed signed provenance metadata in supported file types such as .jpg, adhering to an open standard for recording content provenance. Only Claude models launched on or after August 2 will have the new watermarks; older versions will be updated.

These modifications aim to comply with EU regulations on AI, which require AI-generated content to have machine-readable markings. Non-compliance could lead to significant fines. Anthropic plans to implement these watermarks globally, not just in Europe. The company is among 190 organizations that signed a voluntary EU code of good practice, alongside tech giants like Google, Meta, Microsoft, Mistral, and OpenAI.

Google is developing SynthID, an advanced digital watermarking and detection technology, and is advocating for industry leaders to adopt this standard. Anthropic has not released a watermark detection tool as required by the EU code of good practices.