Anthropic has confirmed it will embed imperceptible watermarks into text generated by Claude models launched in the EU on or after August 2, 2026 and attach digitally signed provenance metadata to generated files, laying out the commitment in a new support page detailing its compliance plan for the EU AI Act’s transparency obligations, which became enforceable August 2, 2026.
The watermarking applies to every Claude model launched in the EU on or after August 2, 2026, at launch, and Anthropic said it is working to extend marking support to models released before that date. The obligation stems from the AI Act’s requirement that providers add machine-readable marks enabling detection of AI-generated or manipulated content. Anthropic signed the EU’s Code of Practice on Transparency of AI-Generated Content as both a model provider and a system provider, joining roughly 190 organizations that signed ahead of the August 2 deadline.
How the Watermark Works
The support page describes two complementary marking techniques. For text, the model weaves an invisible watermark directly into the generated output. Anthropic says it does not alter meaning, quality, or readability. Because the mark is embedded in the text itself rather than attached as metadata, it survives copy-and-paste and may persist through some editing.
“Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing,” the page states. “Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.”
For files, Claude attaches signed provenance metadata following the C2PA open standard, the Coalition for Content Provenance and Authenticity specification used across the industry. The signed label signals that a file was processed by Claude and allows detection of tampering. Supported formats include .svg, .png, and .jpg.
The marks apply across all Claude surfaces: the API platform, the Claude app, Claude Code, Claude Cowork, and Claude Tag. They also apply when supported models are accessed through AWS, Google Cloud, or Microsoft Foundry, though signed provenance metadata may not be supported on every platform depending on available features. Marking applies worldwide, not just in the EU.
What the Marks Cannot Guarantee
Anthropic’s support page is direct about the limitations, a section that reads as much as a legal disclaimer as a technical one. A detected watermark indicates content may have been processed by Claude but does not confirm full provenance. Someone who used Claude to proofread or translate a document would produce output carrying the mark even though the underlying ideas originated elsewhere.
The inverse also holds. Content generated by Claude may not carry a detectable mark if it came from a model released before marking was supported, if the text was heavily edited or paraphrased, if the passage is too short for a reliable signal, or if metadata was stripped through format conversion or screenshots.
Anthropic said it will publish technical documentation on detection mechanisms but has not yet done so. The company committed to supporting users and third parties in detecting Claude’s marks, as the Code of Practice requires, but the detection tooling remains forthcoming.
The Compliance Architecture
The EU’s transparency framework gives companies two paths. Sign the Code of Practice and follow its measures for a streamlined compliance route, or demonstrate compliance through alternative means the Commission considers equivalently adequate. Anthropic chose the code, as did Google, Meta, Microsoft, OpenAI, and several other major AI companies. Unite.AI covered the mandatory labeling obligation when it took effect and Google’s earlier signing of the same code.
Enforcement falls to national market surveillance authorities in each EU member state, with the Commission’s AI Office supervising systems under its own jurisdiction. The AI Office plans to launch two task forces in September 2026 where signatories will share implementation practices and provide feedback on the transparency measures.
For companies building products on Claude, Anthropic’s page includes a pointed note: those deployers should independently assess what the EU law requires of their own products. The company’s watermarking covers its obligations as a provider; it does not automatically satisfy the separate obligations of businesses deploying Claude in their own services. Anthropic said it will share technical guidance to support deployers in meeting their own transparency duties.
The support page says Anthropic will update the article and publish more detailed technical guidance as it becomes available. The September task force meetings will be the first structured venue where signatories compare implementation approaches and surface gaps between the code’s requirements and what companies have actually shipped.
