Anthropic announced on August 11 that Claude models launched from August 2, 2026 onward will embed an invisible watermark in generated text, worldwide, as part of implementing its EU AI Act Article 50 transparency commitments. The misconception spreading already: that this is hidden characters or encrypted tags in your text. It is neither, and I think the mechanics matter for anyone who will be asked to act on a detection result.
Anthropic has not disclosed its algorithm; technical detection documentation is forthcoming for third parties. But the published lineage of this class of watermark points to a common design pattern. Kirchenbauer et al. (ICML 2023) bias token selection toward a pseudorandomly chosen “green list” at each step. DeepMind’s SynthID-Text, published in Nature in 2024, used a related scheme for Gemini outputs, relying on sampling over candidate tokens rather than a simple green-list logit bias. These common ideas are more interesting than the individual implementation: use keyed, context-dependent randomness to slightly influence token selection during generation, so the resulting token sequence carries a statistical fingerprint that a detector can later test against a threshold. Whether Claude uses something like this remains unknown until Anthropic publishes the technical details.
Note what none of these designs involve: public-key cryptography. Text watermarking of this kind is symmetric, a keyed PRF, so verification runs through whoever holds the key. Expect key-gated detection, not a public verifier. Signatures belong to the file side: Anthropic’s C2PA provenance metadata for images is public-key signed and behaves like a credential. The text watermark behaves like a statistical fingerprint.
That difference sets the boundary I would brief a review team on: a statistical mark yields a confidence level, not a proof. Anthropic’s own page says detection shows content “may have been processed by Claude” and can fade under heavy editing, paraphrase, translation, mixing with other writing, or passages that are simply too short. Absence of a detected mark does not prove human origin. One signal among several, weighted like one.
Sources:
Anthropic. “How Claude marks AI-generated content.” https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
Kirchenbauer, Geiping, Wen, Katz, Miers, Goldstein. “A Watermark for Large Language Models.” ICML 2023, PMLR 202:17061-17084. https://proceedings.mlr.press/v202/kirchenbauer23a.html
Dathathri et al. “Scalable watermarking for identifying large language model outputs.” Nature 634, 818-823 (2024). https://www.nature.com/articles/s41586-024-08025-4