Anthropic said future Claude models will embed a hidden statistical marker in the text they generate, so that a passage can later be tested for whether Claude likely wrote it. The change is being made to comply with EU transparency rules on AI-generated content and is being applied worldwide because Anthropic cannot yet limit it by region. The marker carries no information about the user, does not change output quality or price, and works poorly on short passages, code and lightly edited human writing.
What changed
Claude's text output carried no hidden marker indicating it was machine-written.
What it unlocks
Checking, after the fact, whether a passage of text was likely produced by Claude, once the planned detection service is available.
- around 190 signatories to the code
- EU marking rules apply from 2 Aug 2026
- transition period for models before 2 Aug 2026
What you need to act on it
- a future Claude model
- detection requires Anthropic's forthcoming watermark detection API
Sources