Anthropic is introducing an invisible watermark in all texts generated by its Claude AI, according to ARY News. All models released since the beginning of August already have this hidden tracking feature, and the company is also updating older systems ahead of the strict European Union deadline in December.
This week the mark will be included in the outputs of the Claude Mythos 5, Opus 4.7 and Sonnet 4.6 models. The rollout for Opus 4.6 has been postponed until November 16, when older models such as Opus 4.5, Sonnet 4.5 and Haiku 4.5 will also receive the update. Since tracking is added at the model level, users cannot turn it off, regardless of whether they use the official web applications, the API or cloud services such as Amazon Web Services and Google Cloud.
Anthropic uses a modified version of Google's SynthID technology, which slightly changes word choice in order to create a mathematical pattern in the text. According to internal documents, the signature survives basic copy and paste, but heavy editing or translation can break it. The company states that the watermark does not contain users' personal data or conversation history. Ordinary users cannot currently check the mark, because the detection tool is available only in a private preview version for law enforcement, researchers and the media.
ARY News also notes technical shortcomings of the system. For API users, the watermark works reliably on older models only if parameters such as temperature and top_p remain at default values. The system also does not distinguish text that Claude wrote from scratch from a human document that the AI only proofread. Anthropic introduced this global change in order to comply with Article 50 of the new EU artificial intelligence law, which differs from OpenAI, which applies watermarks only for European ChatGPT users and makes them optional for API clients.
What this means for businesses
For companies that use Claude via the API, the change means that generated text will carry a hidden mark that cannot be removed, which may affect internal content review processes. Adjusting model parameters could weaken detection reliability, so firms that depend on verifying text authenticity should take these limitations into account. Since the rule applies globally, not only in the EU, companies operating in multiple markets should align their internal policies on the use of AI tools with the new requirements.
This summary was written with AI based on the cited source and checked by automated rules. For exact details, read the original article.