Anthropic has activated an invisible text watermark in the outputs of any Claude model released on or after August 2. According to the company, the watermark is embedded in the output’s statistical patterns so that it survives copy–paste and some editing. The change has been applied across Anthropic’s five product lines and is enabled globally.
Older Claude models must be brought into compliance by December 2; Anthropic says these models are being retrofitted but has not published a detailed timeline for that work.
What prompted the change?
Anthropic attributes the move to Article 50 of the European Union’s AI Act, which sets requirements for AI systems sold into the EU. Despite that rule targeting the European market, Anthropic turned on the watermarking feature for users worldwide rather than limiting it to Europe.
What is known and what hasn’t been disclosed?
The company describes the watermark as a pattern embedded in the statistical properties of the generated text, but it has not published technical details about how the watermark operates. Paid users are not able to disable the watermark, and Anthropic has not yet released a detection tool for third parties to identify the watermark in model outputs.
Implications and reactions
Observers have noted that applying a regulation-driven measure globally — without a regional pilot or an opt-out for users — raises questions about scope and transparency. Some commentators argue that enabling a non-disableable watermark worldwide could create an opportunity for competitors that refrain from similarly aggressive measures to attract users.
At the same time, compliance with the EU AI Act is a legal consideration for any company operating in the European market, and the lack of a public detector and detailed technical disclosure leaves open questions about accountability, user control, and how robust the watermarking approach will be in practice.
Context
OpenAI is also subject to Article 50 of the EU AI Act but, according to the same reporting, has not announced a matching policy. The interaction between evolving regulation and platform-level technical responses will remain a key issue for developers and users of large language models.
In short: Anthropic has globally enabled an invisible watermark in its Claude outputs to address the EU rule; technical details and a detection tool remain unpublished, and older models are being retrofitted with a December 2 compliance target.
