Overview
- Anthropic began adding the invisible watermark to Claude models released since August 2 and says it will roll the mark out to older models to meet the EU AI Act and the EU transparency code.
- The mark is a statistical fingerprint produced by biasing which near‑equally probable tokens the model selects so a holder of the verification key can detect a pattern rather than any visible label.
- Because Claude generates or regenerates text token by token, the watermark can appear on writing the model only edited or proofread and it can be weakened or removed by heavy rewriting, translation, or very short or code‑like outputs.
- Anthropic has promised detection tooling and a text‑detection API but has not published detectors, verification keys, or empirical error rates yet while developers have already released tools that claim to obfuscate or strip the signal.
- The rollout has prompted some users to cancel subscriptions, stoked worries about false attribution in education, hiring and copyright disputes, and created a fast‑moving adversarial cycle between providers, verifiers, and circumvention developers.