Anthropic explains how Claude’s text watermark works

Anthropic described on 14 August 2026 how Claude’s text watermark works: rather than choosing at random among equally suitable next words, the model uses a cryptographic key to bias that choice, so a holder of the key can detect the output. No user information is embedded.

Why it mattered It is one of the first detailed accounts by a major laboratory of how it meets the EU AI Act’s machine-readable marking rule, which became enforceable on 2 August 2026.