Anthropic is embedding a concealed marker in every passage its Claude AI produces, a step linked to a strict European Union regulatory deadline in December. Every model released since early August already carries the feature, and the company is updating older systems ahead of that date.
This week, outputs from Claude Mythos 5, Opus 4.7 and Sonnet 4.6 will begin carrying the machine-readable mark. Anthropic has pushed the rollout for Opus 4.6 back to November 16, the day Opus 4.5, Sonnet 4.5 and Haiku 4.5 are also set to receive the update.
How the Claude text watermark works
Users cannot switch the tracking off, because it is built into the core model. The access route makes no difference: web apps, API links and outside cloud platforms such as Amazon Web Services and Google Cloud all carry it.
Anthropic adapted Google’s SynthID technology so the marker stays hidden from readers. The system slightly alters word choices to create a mathematical pattern within the text. According to internal documents, the signature survives simple copying and pasting, though heavy editing or translation may disrupt it. The company stressed that the marker holds no personal user data or conversation history.
Detection limits and developer loopholes
Ordinary users currently have no way to check these watermarks. Anthropic has restricted its detection tool to a private preview for law enforcement, researchers and media organizations.
Several technical gaps exist. For API developers, the watermark reliably appears on older models only when temperature and top_p remain at default values, and changing those settings can bypass it. The system also cannot distinguish text Claude wrote from scratch from a human document that the AI only checked for grammar.
Anthropic said the global change complies with Article 50 of the new EU AI Act. The approach differs sharply from OpenAI’s, which applies watermarks only to European ChatGPT users and leaves the feature optional for API customers.
