Policy
Anthropic's Text Watermarks Turn The EU AI Act Into A Product Architecture Test
Anthropic says models released after August 2 will watermark generated text and files to meet European transparency requirements, raising practical questions about persistence, detection and meaningful disclosure.
By Michael C ยท
Anthropic says it will watermark text generated by its models, including Claude, to comply with European transparency rules. According to an updated support page, models released after August 2 will automatically add a watermark to generated text and files, with the company using the C2PA standard for files.
The policy shows how regulation is beginning to change the architecture of AI products rather than just their legal terms. The EU AI Act's Transparency Code requires systems to mark AI-generated or edited material in a machine-readable form. A watermark that follows a copy-and-paste path is an attempt to make that obligation survive ordinary use.
The difficult questions are implementation questions. Anthropic says the mark may persist through some editing, but the threshold at which it can be removed remains unclear. Watermarking is most valuable when an independent system can detect it reliably, users understand its limits, and routine formatting changes do not make the signal disappear.
Anthropic says the measure applies across Claude products and that it will extend support to older models. It joins other companies that have committed to the European code. This is a broad shift from voluntary provenance language toward systems that platforms and regulators can test.
A watermark cannot settle every authorship dispute or prevent deception on its own. It can, however, supply a piece of evidence that was previously missing. The credibility of the approach will depend on whether that evidence is durable enough to help a publisher, platform or reader make a real decision.
Topics: Anthropic, EU AI Act, watermarking