Anthropic details watermarking mechanism for Claude-generated text to comply with EU AI Act

Anthropic published a blog post detailing how it will watermark text generated by its Claude models to comply with the EU AI Act’s Transparency Code, which mandates systems to identify AI-generated content. The company states that watermarks are embedded through patterns in responses that are "undetectable to the reader, but detectable to anyone with a key that encodes it." Anthropic emphasizes that watermarking does not affect output quality and that watermarked responses are indistinguishable from unwatermarked ones to human readers.

The company says it will use the SynthID-Text approach developed by Google DeepMind in 2024, and plans to release a watermark detection API. Anthropic distinguishes its watermarking from AI detection tools like those offered by Pangram, which look for stylistic "tells" in writing rather than checking for a cryptographic watermark.

Anthropic addresses concerns about watermark removal through editing. It says light editing likely won’t remove the watermark, while a complete rewrite that replaces every word can. In the latter case, the company argues, the resulting text may no longer qualify as AI-generated. For text proofread or lightly edited by Claude, detectability depends on how much of the content remains original; if most words are human-authored, there is little for the watermark to attach to.

The company notes that code will generally carry fewer detectable watermarks because functional constraints limit the model’s freedom to choose among equivalent options. Watermarks may appear in comments or areas where arbitrary word choices exist, but their impact on the actual code is described as negligible.

Anthropic states that other major model developers have signed the same Code of Practice and will implement their own watermarks, indicating a broader industry move toward standardized content provenance under the EU AI Act.