Anthropic has announced that future versions of its Claude AI models will embed invisible watermarks in the text they generate. The change is designed to comply with the EU AI Act’s Transparency Code, which took effect on August 2, 2026, and requires AI providers operating in the European market to mark AI-generated content in a way that other systems can detect. Several other major AI providers have signed the same Code of Practice and are implementing their own watermarks.
The watermarking technique works at the token level during text generation. When a language model generates text, many individual word choices are essentially interchangeable — the meaning of a sentence is the same whether it ends with “overcast” or “grey.” Anthropic’s method uses those low-stakes choices to embed a hidden pattern, determined by a cryptographic key, rather than a conventional random number generator. The resulting text reads identically to unwatermarked output, but anyone with access to the key can calculate the statistical probability that the text was Claude-generated. Anthropic emphasizes that the watermark carries no identifying information and cannot be traced to a specific user, organization, or conversation. The technique does not add hidden characters, does not require extra tokens, and does not affect response quality. For file outputs such as images, Claude will use the open C2PA standard — the same metadata format used by camera manufacturers and photo-editing software — to attach a cryptographically signed content credential.
A complete rewrite of watermarked text will eventually erase the pattern, but light editing is unlikely to strip it entirely. Anthropic plans to offer a watermark detection API that developers and platforms can use to check whether text was likely Claude-generated.