Anthropic adds Claude AI watermarks for EU law

Anthropic said it will add invisible watermarks to text produced by its Claude chatbot to comply with European Union artificial intelligence rules.

The American AI company unveiled the feature on Monday in San Francisco. It said the watermark will not be visible to humans and will remain in place even after text is cut and pasted.

Anthropic said the mark has limits: detecting a Claude watermark means text may have been processed by Claude, but does not prove Claude originally generated it. The company said people may use Claude to proofread, translate, summarise or convert files, meaning the output can carry a Claude mark even when the underlying material came from elsewhere.

The company also said very short or heavily edited text may not carry a watermark.

When Claude generates a supported file type such as .jpg, Anthropic said it will attach signed provenance metadata in line with an open standard used to record information about content provenance.

Anthropic said only Claude models launched on or after August 2 will include the new watermarks at first, and earlier models will be upgraded. The company said it will use the watermarks worldwide, not only in Europe.

The changes are meant to comply with new EU rules requiring AI-generated content to include machine-readable markings. Anthropic said violations risk drawing stiff fines.

Anthropic was among about 190 organisations that signed a voluntary EU code of good practice late last month, alongside companies including Google, Meta, Microsoft, Mistral and OpenAI. Google has developed SynthID, a non-public digital watermarking and detection technology, and is pushing other major technology companies to adopt it.

Anthropic has not yet released a watermark detection tool called for in the EU code of good practice.