- Anthropic is adding an invisible watermark to Claude's text because the EU AI Act requires it; newer models have it from launch, older ones get it in stages.
- According to The Decoder, on Opus 4.6 and Sonnet 4.6 the mark only applies at default randomness settings.
In short: Anthropic is adding a hidden watermark to text written by Claude. It is an invisible pattern in word choice that a special tool can detect. The reason is the EU AI Act, which requires AI-generated content to be marked. According to The Decoder (31 Oct 2026), on two older models the mark can be lost with a simple settings change.
What Anthropic says
- On its help page Anthropic says it has signed the Code of Practice for Article 50(2) of the AI Act on transparency of AI-generated content.
- Models launched in the EU on or after 2 Aug 2026 support marking from launch, and the mark is still being added to models released earlier.
- The table on the page lists the text watermark for, among others, Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5, as well as Opus 4.6 and Sonnet 4.6. The mark is applied at the level of the model itself.
- Detection is for now in a private preview for regulators, law enforcement, media, fact-checkers, researchers and educational organisations. The free Claude Content Checker only reads C2PA labels in files, not text.
Where the loophole is
The Decoder quotes an Anthropic email to customers. It says that on Opus 4.6 and Sonnet 4.6 the mark only applies at default values of temperature and two similar settings (top_p, top_k). Changing any of them switches it off, and the company advises anyone relying on the mark for compliance to keep the defaults. The same email reportedly says the mark does not change prices or request formats and holds no information about the user. I have not seen the email myself; I rely on The Decoder’s account.
How to read this
The mark is not an AI detector for everyone: for now only selected institutions can check it. Critics cited by The Decoder ask whether choosing words for the mark lowers text quality, and there is no independent data on that yet. I write more about AI Act duties in the AI Act and law hub.