Anthropic opens Claude watermark checks to regulators and the press
Anthropic is giving outside organisations access to an API that checks whether text carries the invisible watermark its Claude models embed. TNW reports that the Claude watermark detection API has shipped in private preview for regulators, media, fact-checkers and other eligible groups. That is narrower than the free public detection API TNW had reported in August. Anthropic says it plans to expand access over time without giving a date. The announcement came with Claude Fable 5.1 and Mythos 5.1.
Artificial Intelligence··Morning
Anthropic puts Claude watermark checks in private preview
Anthropic is giving outside organisations access to an API that checks whether text carries the invisible watermark its Claude models embed. TNW reports that the API has shipped in private preview for eligible organisations, including regulators, law enforcement, media, fact-checkers, independent researchers, educational bodies and EU civil-society groups. Enterprises that have their own compliance obligations are also able to apply. Access is requested through a form. That is narrower than the free public detection API TNW had reported in August. Anthropic says it plans to expand access over time without giving a date. The announcement came with the release of Claude Fable 5.1 and Mythos 5.1.[1], [2]
The mark follows SynthID-Text and the EU AI Act date
The watermark is a variant of the SynthID-Text method Google DeepMind published in Nature in 2024. The method steers the randomness in word selection so that the choices follow a pattern detectable with a key. TNW traces the watermark to the EU Code of Practice on transparency of AI-generated content, which Anthropic signed in July alongside about 190 other signatories. Article 50 of the AI Act has required providers to mark synthetic output in a machine-readable, detectable form since 2 August 2026. The mark applies to Claude models released after that date. It is invisible without the detection tool, and according to Anthropic carries nothing about the user.[1], [2]
Anthropic says light edits survive and a rewrite erases the mark
Anthropic says the mark survives light editing but is erased by a full rewrite. The company says it works poorly on short or fact-dense passages where few alternative wordings exist. It also says the check cannot separate Claude authorship from heavy human editing. Claude models released after the EU transparency requirement took effect carry the mark. The key is what makes that pattern detectable to an outside checker. Those limits are Anthropic's own description of the detector.[1]
