Anthropic is integrating an invisible watermark into Claude’s outputs to support transparency under the EU’s AI Act, marking a significant step in traceable AI technology despite current limitations in proving authorship.
Anthropic is beginning to add an invisible watermark to Claude’s outputs as regulators sharpen their focus on machine-generated content. The company says the system is meant to support transparency obligations under the European Union’s AI Act, which requires synthetic text, images, audio and video to be machine-readable and detectable as AI-generated. Anthropic says it is applying the change broadly rather than only inside the EU because it does not yet have a reliable way to limit the feature by region.
For text, the watermark is not a visible label and does not insert extra characters. Instead, Anthropic says Claude is made to choose between acceptable word options in a way that creates a statistical pattern. In practice, that means the model’s output still reads normally, but a detection tool with the right key can assess whether the text likely came from Claude. The company says this should have negligible impact on speed and no additional serving cost.
Anthropic is also stressing the limits of the approach. A watermark can suggest that Claude helped produce text, but it cannot prove authorship. That matters when users ask the model to translate, proofread or lightly edit existing material, because the final result may contain too little Claude-generated language for the pattern to be clear. The same problem applies to code, where exact syntax leaves less room for the model to vary its word choices. Heavy rewriting can also weaken or remove the watermark, while small edits may not.
For files, Anthropic says it will use a different method. Supported image formats such as PNG, JPG and SVG will carry cryptographically signed provenance data through the C2PA standard, allowing recipients to check whether Claude was involved in creating or processing the file. The company also plans to release a detection API later, which would let third parties test whether text appears to contain Claude’s watermark. The move adds to a wider industry shift towards traceable AI output, but it also underlines a basic limitation: watermarking can improve transparency without making AI detection certain.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
