Claude Text Watermarking is a cryptographic signal 📝Anthropic embeds in Claude's word choices so that a party holding the key can verify Claude likely wrote or edited a passage, without changing how that passage reads.
Announced August 14, 2026, it applies to all 📝Claude generations — including older models — and carries no user, organizational, or conversation data. It operates on low-stakes choices where several wordings are equally valid: rather than sampling randomly, Claude selects among semantically equivalent options using a cryptographic key seeded by the preceding words, so a checker holding that key can test whether a sequence matches the choices Claude would have made. The method derives from Google DeepMind's SynthID-Text, published in Nature in 2024 and traceable to a 2022 proposal by Scott Aaronson. It produces no extra tokens, adds no cost or latency, and showed no statistically significant quality difference in DeepMind's evaluations.
Reliability rises with passage length and falls wherever Claude has few free choices — factual statements with one correct answer, proofreading of human-written text, and code requiring exact syntax. The signal is one-directional: it cannot separate text Claude wrote from text Claude edited, cannot establish that a passage is human-written, and does not detect other models. Light editing may leave it intact; rewriting every word removes it.
Anthropic shipped it to meet the EU Code of Practice on Transparency of AI-Generated Content, signed in July 2026 by roughly 190 organizations under the EU AI Act, and applied it globally rather than gating by region. Image and file outputs instead carry C2PA content credentials. A detection API is announced but not yet released.