Anthropic is introducing digital watermarks for text generated or processed by its artificial intelligence models, including Claude chatbots.
The company will also add metadata to certain files created or processed by Claude, allowing users and third parties to identify AI involvement.
The move is part of Anthropic’s effort to comply with transparency requirements under the European Union’s AI Act.
The EU rules require AI-generated or modified content to carry clear signals indicating its origin.
Anthropic said the measures are designed to provide people with more information about the content they encounter online.
How Claude’s watermarking works
Claude models launched on or after Aug. 2 will include machine-readable marks in text and files generated or processed by the models.
Anthropic said an “imperceptible” watermark will be embedded directly into the text through its backend systems.
Because the watermark is incorporated into the text itself, it can travel with content when users copy and paste it elsewhere.
The company said the mark may also survive certain types of editing.
Watermarking will operate at the model level, meaning it will apply regardless of which Claude product or interface is used.
The system will also cover content that was originally created by humans but later processed by Claude.
For example, text submitted to Claude for proofreading, translation or summarization could receive the watermark.
Anthropic said the watermark may remain even after users modify the processed text or combine it with other material.
However, the company stressed that a detected watermark would not prove that an entire piece of content was generated by AI.
A human-written document could carry the mark simply because it was processed through Claude.
Metadata added to AI-generated files
Anthropic is also introducing signed provenance metadata for files such as JPG images created or processed by Claude.
The metadata will indicate whether Claude created or processed the file and whether the content has been altered.
The company said it is developing detection tools that will allow users and third parties to identify Claude watermarks and provenance metadata.
Anthropic plans to provide more information about those detection mechanisms in the future.
The company is also working to bring the watermarking system to older Claude models.
At the same time, Anthropic warned that the absence of a detected watermark does not prove that content was not created or processed by AI.
For images, metadata can potentially be removed through screenshots, file conversions or other processes.
The development comes as major technology companies face growing pressure to improve transparency around AI-generated content.
Anthropic is among more than 200 companies, including OpenAI, Meta, Google and Microsoft, that have signed the EU’s voluntary code of conduct on AI transparency.
Spotify has also announced plans to label AI-generated artist personas and exclude them from personalized music recommendations.
OpenAI recently expanded its content provenance technology to cover AI-generated audio in addition to images.
Canada is also developing its own approach to AI transparency. The federal government is currently seeking public input on potential rules through Sept. 23.