Anthropic's Claude gets invisible watermarks for EU AI Act compliance

Anthropic has pledged to watermark all text and images generated by its Claude models, announcing the move on a support page aimed at satisfying the European Union’s AI Act. The transparency rules took effect August 2, 2026, with a four-month grace period, so the company will apply machine-readable labels to new models from day one while retrofitting existing ones remains a work in progress. For images, Claude will attach C2PA provenance metadata, the standard already used by Adobe, OpenAI, and Google. For text, Anthropic says it will weave an “imperceptible watermark” directly into responses, without changing meaning, quality, or readability. The watermark travels with copied and pasted text and may survive some editing, and it will be present across every Claude surface—API, Claude Code, Claude Cowork, and Claude Tag—including when models run through AWS, Google Cloud, or Microsoft Foundry.

Despite the ambitious scope, Anthropic concedes the marking systems are not infallible. C2PA metadata can be stripped accidentally when files are uploaded to online platforms, and the company hasn’t yet explained how robust its text watermark is. It is developing detection tools and will publish technical documentation later, though external detectors like Google’s Gemini may not be compatible with Claude content. The practical benefit is real: platforms and users gain a standardized way to identify AI-generated text and images, helping curb synthetic media spread. But because watermarks are invisible and any content lacking marks could still come from AI, these labels serve more as a transparency signal than a definitive filter. Success depends on whether Claude-generated content can be reliably identified across the messy, edited, platform-processed web.