Claude Adds EU-Mandated Text Watermarks Without Extra Tokens

EU rules will shape Claude’s word choices

Anthropic says future Claude models will place a statistical watermark in generated text to comply with the EU AI Act and the transparency code it signed in July 2026. The company plans to apply the mechanism globally at launch because it does not yet have a durable way to scope it by region. Roughly 190 organizations, including other major model providers, have signed the same code.

“We’re implementing watermarking to comply with the EU AI Act. Other major model developers have signed the same Code of Practice and will also be implementing watermarking.” — Anthropic

Anthropic’s account is its official outlet for Claude products, safety research, and policy positions. Its technical explanation says the watermark does not insert hidden characters or an extra notice. When several next-word choices are similarly valid, the model uses a secret key and preceding words to influence the random selection. A detector holding that key can later test whether the sequence is statistically consistent with Claude’s process.

A provenance signal, not an authorship verdict

The implementation is based on the SynthID-Text family described by Google DeepMind in a 2024 Nature paper. Anthropic says internal testing found no practical change in content, creativity, or readability. Because the method generates no extra tokens, it should not change serving price and should have negligible speed impact. The signal contains no information identifying a person, organization, or chat.

Its limits are substantial. Short passages offer too few choices for reliable detection. Factual prose, proofreading, and exact code also leave less room for a watermark. Heavy rewriting can remove the pattern, and a positive result can only indicate that Claude was probably involved; it cannot prove human authorship, legal ownership, or which person used the model. Anthropic says a detection API is coming, while older Claude models will receive watermarking over the coming months under the law’s transition period.

Watch for published false-positive and false-negative rates, detector access rules, interoperability across providers, and how the mechanism survives translation or editing. The source post is on X, and Anthropic’s watermark explainer details the method and its limitations.