How Claude’s text watermark works

Details

As of August 2, 2026, the EU AI Act requires AI providers serving European markets to mark AI-generated content. Anthropic, alongside other major model developers who signed the same Code of Practice, is implementing text watermarking in all future Claude models. The technique works by influencing low-stakes word choices—cases where multiple words carry nearly identical meaning—using a cryptographic key instead of pure randomness. The sequence of these choices becomes statistically consistent with the key, allowing detection by anyone holding it. Crucially, the pattern is invisible to readers: no hidden characters are added, nothing changes about how text reads, and no extra tokens are required. Anthropic is explicit that the watermark carries no identifying information traceable to a specific user, organization, or conversation. Detection yields a probability estimate, not a definitive verdict. The watermark is also not Claude-specific—it produces a statistical signal, not a model fingerprint. This is an industry-wide regulatory response. Multiple major labs have signed the same EU code of practice and will ship their own watermarks. The article does not address interoperability between different providers' detection systems.