Claude texts carry a hidden pattern that can point to AI use

When Claude writes a text, a hidden trace may sit in the word choice itself. Anthropic uses a text watermark that can't be seen with the naked eye, and according to an Italian tech blog, it will be extended to several of the company's older models starting November 3, 2026.

A pattern in the words, not hidden characters

The watermark doesn't consist of invisible characters tucked into the text. It builds on a version of SynthID-Text from Google DeepMind. A language model often chooses between several equally good words. Here, the choice is steered by a key and by the preceding words. That leaves a statistical pattern that can be detected if you have the key.

Anthropic says the marking doesn't degrade quality, doesn't make Claude slower or more expensive, and doesn't contain any information about the user.

What changes on November 3?

According to the blog, which cites an email from Anthropic to organizations, the marking will begin to apply to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. It happens first on Claude Platform, Anthropic's API (the technical connection companies use to build on Claude). On Amazon Bedrock, it may take a few extra days, according to the blog. No dates have been given for Google Cloud or Microsoft Foundry.

The same technique is already in use on newer models such as Opus 5.5, according to the blog. Anthropic's own support page already lists a long range of models as supported. It isn't clear which models are enabled when, and on which platforms.

Background: EU rules

Anthropic has signed the EU code of practice on transparency for AI-generated content, tied to Article 50(2) of the AI Act. New Claude models launched in the EU from August 2, 2026 support machine-readable marking from the start. Anthropic is working to add marking to older models. The marking is listed across Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag, but some platforms or features don't necessarily support all marking types.

Images and other supported files instead get signed C2PA content credentials, meaning metadata about their origin. These can disappear, however, during format conversion or when someone takes screenshots.

Limitations and who can check

Anthropic says the watermark can't distinguish between text Claude has written and text Claude has heavily edited. It works poorly on short texts, on factual texts with few options, and on code that requires exact output. If the entire text is rewritten, the traces can disappear.

According to Anthropic, a missing mark therefore doesn't prove that the content wasn't made by AI. A detected mark only means that Claude may have processed the content.

A detection API is in private preview. It is reserved for approved organizations such as EU authorities, police, media, fact-checkers, researchers, educational institutions and civil society, as well as companies with compliance obligations. Anthropic plans broader access, but ordinary users can't check text themselves yet.