Anthropic will now embed invisible identifiers into every text response and image produced by its Claude models, applying the change globally rather than limiting it to European users. The move follows the company’s endorsement of a European Union transparency framework that took effect on August 2, 2026.
The San Francisco AI developer confirmed through a support document that it has signed the EU AI Act’s Article 50(2) Code of Practice covering transparency for AI-generated content. Anthropic joins as both a provider of generative models and the systems built on them.
This positions Anthropic among the first major AI companies to spell out operationally how it plans to comply with rules that will eventually mandate identification for most synthetic media circulating within the EU. The obligations split responsibility across two camps. Providers building the models must ensure outputs can be recognized as artificial. Deployers, the much larger group of organizations using those models to publish material, bear separate disclosure duties.
For Claude’s text output, Anthropic embeds an imperceptible watermark that survives copy-paste actions and light editing without altering readability or meaning. For supported image formats including PNG, JPG, and SVG, the company attaches signed provenance metadata following the C2PA open standard. That framework, already adopted by Adobe, the BBC, and multiple camera manufacturers, records a file’s origin and flags any tampering.
Models released in the EU on or after August 2, 2026 carry the marking from launch. Older models receive updates to support it. The system spans Anthropic’s API, consumer apps, Claude Code, and enterprise deployments. A key caveat: a detected mark only signals that content “may have been processed by Claude.” Its absence does not prove human authorship, since heavily edited text or older model output may lack any identifier.
By the numbers, roughly 190 organizations have signed the EU code. About 82 signed the provider section while 152 signed the deployer portion. Provider signatories include Google, Meta, Microsoft, OpenAI, and Synthesia. Deployers include Getty Images, Lenovo, and Lufthansa.
Anthropic’s emphasis on text watermarking sets it apart. Text remains the hardest format to mark reliably, since paraphrasing or retyping can strip embedded signals. That technical hurdle explains the law’s “as far as technically feasible” qualifier and Anthropic’s framing of its marks as indicative rather than conclusive.
Enforcement remains nascent. Non-compliance penalties could eventually reach €15 million or 3 percent of global annual turnover, whichever is higher. Regulators hold inspection powers over models, though enforcement bodies are still building capacity. In the meantime, businesses using Claude and rival systems face a practical question: which AI-assisted outputs need labels under Article 50, and how to attach them before enforcement accelerates.
