Anthropic has started marking the text written by its newest Claude models with a statistical watermark, to meet the machine-readable marking duty of the EU AI Act’s Article 50, which has applied since August 2, 2026. Its explainer, updated on September 1, describes how the mark works and, just as usefully, what it cannot do.
How it works
The method is a version of SynthID-Text, which Google DeepMind published in Nature in 2024. When a model has several equally reasonable words to choose from, the watermark changes only the source of randomness behind that choice. The resulting pattern is invisible to readers and detectable by anyone who holds the key. Nothing is added to the text, there are no hidden characters, no extra tokens and no price change. Anthropic says quality is unaffected, citing its internal tests and Google’s published results on its own system.
The mark carries no identifying information: it cannot be traced to a user, an organization or a conversation, and it does not change who owns the output.
Which models and surfaces
According to Anthropic’s Help Center, Claude models launched on or after August 2 support the marking, and it names Fable 5.1 and Mythos 5.1 as currently supported. Because it is applied at the model level, the mark follows the output across the API, the Claude apps, Claude Code and the cloud platforms, and Anthropic applies it worldwide because it does not yet have a durable way to limit it by region. Models launched before August 2 fall under the EU transition period, and Anthropic says it will add marking to them over the coming months, without naming models or dates.
Files are handled differently: supported images carry a signed C2PA content credential in their metadata, which a free Claude Content Checker can read.
Who can detect it
Detection is in private preview. Access goes to the groups EU law names, such as regulators, law enforcement, media, fact-checkers, researchers and civil-society organizations, and to companies that need to verify the mark for their own compliance. Anthropic says it plans to widen access over time.
What it cannot prove
Anthropic lists the limits itself. A detected mark indicates likelihood, not proof. It cannot confirm that a person wrote a text, cannot identify another AI, and cannot tell “Claude wrote this” from “Claude heavily edited this.” A missing mark does not mean the text was written by a person.
The signal is stronger in long texts and weaker where there are few equivalent word choices: factual passages, proofreading and lightly edited human writing. Where an exact output is required, as in code, the mark is not applied. Light editing will probably not remove it, but a complete rewrite will. Anthropic has published no detection rates or minimum-length thresholds for its own deployment.
For companies building on Claude, the Help Center adds a practical note: each deployer has to assess independently what Article 50 requires of it.
Related reading
- How Claude marks AI-generated content — Claude Help Center
- Scalable watermarking for identifying large language model outputs — Nature, 2024
- What the AI Act demands from August 2 — IA al Día