Claude will begin digitally watermarking marking AI-generated text and images — Anthropic details how it'll comply with the EU's Artificial Intelligence Act
European users of Anthropic's Claude models can expect that future versions of those tools will begin embedding digital marks in generated text and images to identify them as the product of AI.
Claude, an AI model, will soon digitally watermark both text and images it generates, as per the European Union's Artificial Intelligence Act (AIA). This move aligns with the law's Article 50, which mandates AI service providers to implement identifiability measures by August 2, 2026. Anthropic has released a guidance article detailing how new versions of Claude will adhere to these requirements.
Unlike image watermarking, which is more complex, Claude will subtly bias the selection of tokens (parts of words) during text generation. When analyzed, the text will exhibit a specific statistical pattern, signifying the presence of a watermark. Anthropic plans to offer detection tools to identify these patterns in subsequent technical documentation.
The watermarking process won't impact the quality and meaning of the generated text. For images, Claude will attach provenance information using the Content Authentication and Provenance Alliance (C2PA) standard. This digital certificate will be invalidated if the image is altered, ensuring the authenticity of the content.
However, Anthropic notes that text slices under 200 tokens won't be watermarked due to insufficient data for reliable watermarking. As for images, while metadata will be added, actual image watermarking isn't mentioned, though it's required by the Code of Practice. This could change if Anthropic decides to implement such a feature.
To comply with the AIA, Anthropic is also working to add these capabilities to existing Claude models by December 2, 2026. The guidance begins by discussing models launched in the EU but clarifies that watermarking will apply to outputs from supported models worldwide. It's worth noting that direct quotes in responses might be erroneously watermarked, and the absence of a text watermark doesn't indicate non-AI generation.
Lastly, the AIA requires service providers to provide tools for detecting these watermarks, and Anthropic promises to share more details in forthcoming documentation.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
