In a move designed to meet the strict transparency requirements of the European Union’s AI Act, Anthropic — the developer of the popular Claude family of large language models — has rolled out mandatory invisible text watermarks and digital metadata labels for all AI-generated content produced by its systems across the globe. The new rule, which went into effect for all new Claude models launched in the EU after August 2, 2026, will eventually be extended to older versions of the model that were rolled out before that date, covering every major platform where Claude is available: the official Claude web interface, Claude API, Claude Code development tool, and the Claude Cowork collaboration platform. Unlike optional watermarking systems that let users choose whether to add identification markers, Anthropic has confirmed that no opt-out is available for any users of its services, according to reporting from Forbes.
The watermarking system is engineered to be unobtrusive to human readers: the embedded markers are imperceptible to the naked eye, persist through standard copy-and-paste workflows, and do not alter the readability of generated text. For digital image outputs in common formats such as PNG and SVG, the system adds hidden metadata tags that track the content’s origin and enable detection of unauthorized tampering. Anthropic has also sought to ease user concerns by clarifying that the presence of a watermark only confirms that Claude interacted with the content, not that the entire work was fully AI-generated. The company added that watermarks can be easily removed through common content modification processes, including heavy editing, paraphrasing, cross-language translation, and screen capture of text. Moving forward, Anthropic plans to publish open technical guidance to help third-party developers integrate watermark detection tools into their own workflows, so that organizations can meet their own transparency obligations under the EU AI Act.
Despite these clarifications, the policy has sparked immediate and widespread pushback from users across major social platforms including X (formerly Twitter) and Reddit, with the sharpest criticism coming from professional groups that rely on Claude as a productivity tool: software developers, freelance writers, and editorial professionals. Many users have raised concerns that even minor use of Claude for routine tasks — such as proofreading a human-written manuscript, debugging a section of custom code, or converting a file between formats — will leave a watermark on the final work, leading to incorrect labeling of original human-created content as fully AI-generated. Conservative radio host Erick Erickson was among the high-profile critics, publicly calling the policy “ridiculous” and noting that even his own original writing would end up carrying an AI watermark after receiving basic editing assistance from Claude.