Anthropic, the company behind Claude AI models, has introduced text watermarks and digital file labels to Claude outputs globally to comply with the European Union AI Act. Models launched in the EU after 2 August 2026 now include machine-readable markings, with plans to add them to earlier models.
The policy covers outputs across the Claude API, main web platform, Claude Code, and Claude Cowork.
The decision sparked quick criticism on X and Reddit, particularly among software developers and editors. Users worry that original human writing will be flagged as AI-generated after using Claude for proofreading, coding, or converting files. Radio host Erick Erickson labelled the change ridiculous, saying his original writing could carry a watermark after getting basic help.
Generated text now includes an imperceptible watermark that survives copying and pasting while keeping the output easy to read. Supported file formats, like PNG and SVG, are tagged with metadata to record where files come from and detect tampering. According to Forbes, users cannot opt out of watermarking.
Anthropic noted that a detected mark shows Claude touched the content but does not prove who originally wrote it. Watermarks can also be lost through heavy editing, paraphrasing, translation, or screenshots. The company plans to release technical guidance to help developers detect these marks and meet their EU transparency obligations.
