Anthropic confirmed on August 11 that it has begun embedding invisible watermarks directly into the text generated by its Claude AI models. All Claude models released from August 2, 2026 onward automatically include machine-readable marking in both generated text and files, along with signed provenance metadata for images and other outputs.
The move is designed to comply with Article 50 of the EU AI Act's transparency rules, which took full effect on August 2, 2026. Under these rules, AI providers must make it clearer when content is AI-generated. Fines for non-compliance can reach up to €15 million or 3% of worldwide annual turnover.
Critically, Anthropic is applying the watermarking globally — not just within the European Union. The company said the decision to roll it out worldwide rather than region-specific was driven by the complexity of managing different versions across jurisdictions.
The watermark is designed to be invisible to human readers but detectable by machine tools, allowing automated systems to identify AI-generated content. Anthropic's approach adds to a growing industry trend: Google already watermarks Gemini's output, and the EU AI Act is pushing all AI providers toward similar transparency measures.
This is one of the first major compliance milestones under the EU AI Act's broader enforcement timeline, which will progressively introduce requirements for high-risk AI systems, biometric identification, and foundation models through 2027 and 2028.
