Anthropic has confirmed that new Claude models will include an invisible statistical signal in generated text, introducing a new approach to AI transparency as the EU AI Act’s Article 50 requirements take effect.
The marking will apply to Claude models launched from August 2, 2026, and Anthropic says the system will be used globally across its API, Claude app, Claude Code, Claude Cowork and Claude Tag, including deployments through AWS, Google Cloud and Microsoft Foundry.
Unlike a visible label or metadata tag, the signal is embedded statistically within the sequence of words generated by Claude. The system subtly influences word selection according to a hidden pattern, allowing specialised detection tools to identify text that has passed through a supported Claude model.
Anthropic says the approach is designed to survive ordinary copying and pasting and may remain detectable after some light editing. However, the company has not yet published the underlying algorithm or released a public detection tool, leaving its effectiveness in real-world use yet to be established.
The watermark is also not intended to prove authorship. A detected signal would indicate that Claude processed the text at some point, but would not establish who originally created the ideas or content. Similarly, the absence of a detectable signal would not prove that text was written entirely by a human.
Detection may also become more difficult after extensive rewriting, translation, mixing AI-generated and human-written material, or when analysing very short passages.
Anthropic plans to handle generated files differently. Supported image formats such as PNG, JPG and SVG will receive digitally signed provenance metadata based on the C2PA standard. Unlike the statistical text signal, this metadata is stored separately from the file content and can potentially be removed through actions such as screenshots, format conversions or resaving files.
The company has described the measures as transparency tools rather than definitive AI-authorship tests, with external detection tools still under development.
