Anthropic has outlined a Claude AI content watermark program for supported models, saying generated text will contain an imperceptible, machine-readable signal. The company has not disclosed how that signal is encoded or how a detector will find it, despite presenting the program as part of its commitments under the European Union’s AI transparency code.
In its support documentation, Anthropic says it signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. The company says it will update the page with more detailed technical guidance as it becomes available.
That leaves the important engineering question unanswered. Anthropic describes the text mark’s intended behavior, but not the mechanism that places it in output or the method a third party would use to identify it.
How will Claude’s AI content watermark work?
Anthropic says a supported Claude model will weave an imperceptible watermark into generated plain text. According to the company, the mark will not alter the response’s meaning, quality or readability. It says the watermark should travel when text is copied and pasted and may remain detectable after some editing.
The company says marking occurs at the model level, so it is intended to apply regardless of which Claude product or surface produced the text. The Hindu reported that the program applies globally across supported models used through Claude, its API, Claude Code, Claude Cowork and Claude Tag, as well as supported access through AWS, Google Cloud and Microsoft Foundry. The report also said not every platform or feature will support every kind of mark.
Anthropic uses a different approach for supported generated files. Rather than the text watermark, those files can receive digitally signed provenance metadata where supported. The Hindu identified SVG, PNG and JPG among the file types covered. Provenance metadata is a signed record attached to a file that can indicate that Claude processed it and whether the record was altered. It is separate from the company’s claimed text-level watermark.
What can a Claude mark prove?
Less than some users or platforms may hope. Anthropic says a positive detection result means content may have been processed by Claude. It does not establish that Claude originated the work or wrote all of it. A person could bring human-created material to Claude for editing, translation, summarization or another form of processing, and the result could still carry a mark.
Absence is no clean verdict either. The Hindu reported that Anthropic warned that extensive editing, short passages and stripped file metadata can produce false negatives. Older models are also in a transition period, according to the report.
- Known: Anthropic says supported text will receive an imperceptible watermark, while supported files can carry signed provenance metadata.
- Not disclosed: The company has not published the text embedding technique or detection mechanism in the material available so far.
- Not proof of authorship: A detected mark signals possible Claude processing, not complete or original Claude authorship.
Anthropic says it is working on detection support for users and third parties and plans to publish technical documentation. The engineering details and detection documentation remain forthcoming.
This story draws on original reporting from Daring Fireball.
