Anthropic’s Claude Watermarks: How Invisible Text Marking and C2PA File Signatures Work Under the EU AI Act

Anthropic uses different technical approaches for text and for generated files .

Text watermarking. When a supported Claude model generates text, it weaves an imperceptible signal directly into the token sequence . The watermark is invisible to the human eye during normal reading and does not change the meaning, quality, or readability of the response . The mark travels with the words when users copy and paste text elsewhere, and it is designed to be detectable by machines using a specialized detection tool .

File provenance metadata. For generated file types—specifically .svg, .png, and .jpg—Claude attaches signed C2PA (Coalition for Content Provenance and Authenticity) provenance metadata . This cryptographic signature is embedded in the file's metadata and can signal whether the file was processed by Claude and whether it has been tampered with .

Scope of the Rollout: Global, Not Just EU

Although the watermarking initiative is a direct response to EU requirements, Anthropic is applying it worldwide. "Marking will apply to output from supported models wherever Claude is offered, globally," the company stated .

• All Claude models launched on or after August 2, 2026 support marking at launch .
• Models released before that date are being retrofitted to support marking .
• Anthropic also formalized its commitment by signing the Code of Practice under Article 56 of the EU AI Act .

Real Limitations: Why Detection Is Not Conclusive

Anthropic has explicitly stated that these markings are not conclusive proof of AI origin . Several fundamental limitations reduce their reliability:

• Text watermarks can be stripped or degraded. Paraphrasing, rewriting, translation, or other manipulations can destroy or weaken the watermark .
• C2PA metadata can be removed. File metadata can be stripped, re-saved, or edited, breaking the provenance chain entirely .
• Detection is probabilistic, not deterministic. The watermarking approach means some AI-generated text may go undetected, and some human text could theoretically trigger a false match.
• No public detection tool has been announced. While the watermark is machine-readable, it remains unclear how widely available the detection infrastructure will be for third parties.

Connection to the EU AI Act Article 50

The rollout is a direct compliance response to Article 50 of the EU AI Act, whose transparency obligations became enforceable on August 2, 2026 .

What Article 50 requires:

Providers of generative AI systems must ensure their outputs are marked with "effective, reliable, robust and interoperable machine-readable marks" that allow outputs to be detected as AI-generated or AI-manipulated . Certain AI-generated or manipulated content must also carry clear and visible labels for end users .

The rules apply to any AI system placed on the market or put into service in the EU, regardless of where the provider is based—giving the Act extraterritorial effect . Non-compliance can result in fines of up to €15 million or 3% of worldwide annual turnover .

The European Commission published official guidelines on July 20, 2026, and confirmed the Code of Practice on Transparency of AI-Generated Content as an adequate means of demonstrating compliance . Even though adherence to the code is voluntary, the transparency requirements under Article 50 are legal obligations .

Anthropic's compliance approach:

Anthropic's global rollout means Claude users everywhere—not just in Europe—will see these markings . The company signed the EU's Code of Practice on Transparency, which serves as a voluntary compliance pathway under Article 56 . New Claude models support marking at launch, and older models are being retrofitted .

What This Means for Users and the Industry

For anyone who uses Claude or encounters its output, these watermarks add a layer of traceability—but not certainty. The practical reality is that a determined user can strip either the text watermark or the file metadata with simple post-processing steps. The markings are most useful as a deterrent and as a baseline compliance measure: they make casual AI-generated content easier to identify without disrupting the normal reading experience.

The broader industry significance is that Anthropic's move is one of the first large-scale deployments of mandatory AI content transparency under the EU AI Act. Other providers will need to implement similar measures, and the technical and practical limitations Anthropic has acknowledged will likely become central to the ongoing debate about how enforceable—and how useful—these transparency rules really are.