Anthropic plans to watermark all AI output from its assistant Claude going forward. With this, the company is responding to the requirements of the EU AI Act. Specifically, Anthropic has signed the Code of Practice for Article 50(2), committing to mark AI-generated content in a machine-readable way.
Watermarks and signed metadata
New models placed on the EU market since August 2, 2026, support the markings from the start. The labeling applies worldwide and across platforms, meaning Claude Platform, Claude, Claude Code, Claude Cowork, and Claude Tag. Models released before that cutoff fall into a transition period. Anthropic is working on retrofitting them but hasn't given a concrete timeline, and it acknowledges that not every platform or feature supports all marking types equally well.
Technically, Claude uses two complementary methods. First, invisible watermarks are embedded directly into the generated text. They're meant to survive when the text is copied or edited afterward. Second, for supported file types like SVG, PNG, or JPG, Claude attaches signed metadata following the C2PA standard. Such a label signals that the file was processed by Claude.
Why the marking isn't proof
Anthropic stresses that the machine-readable labels carry no evidentiary weight. A marking only indicates that Claude was involved in the content, it doesn't document the full history of how the material came about. That matters because Claude often serves as a proofreader, translator, summarizer, or converter. The result then carries a marking, but the underlying text or file may well be human-made. That could lead to false suspicions.
Conversely, a missing marking doesn't automatically mean there's no AI content. Texts can be heavily revised, reworded, or so short that a reliable signal is no longer possible. According to Anthropic, companies that build Claude into their own products should check for themselves what Article 50 means for their offering. The approach itself isn't new: Google's SynthID labels AI-generated text in a similar way, and ChatGPT temporarily left such markers behind too.