Anthropic will begin embedding invisible watermarks in text produced by its newest Claude models, a move driven by the European Union’s AI transparency rules that will nonetheless apply globally. The company also plans to attach signed provenance data to certain file formats, giving users a way to trace content back to Claude.
The policy kicks in for any Claude model released on or after August 2, 2026. While Article 50 of the EU AI Act prompted the change, Anthropic confirmed the markings will appear wherever supported models operate, not just within European borders.
Anthropic has joined the Article 50(2) Code of Practice on Transparency of AI-Generated Content as a provider of both generative AI models and systems. The company’s approach relies on two machine-readable methods: one watermark woven directly into generated text, and another metadata layer attached to files.
Outputs will carry markings across Anthropic’s full product lineup, including the Claude Platform API, Claude, Claude Code, Claude Cowork, and Claude Tag. Embedded watermarks also appear when users access supported models through AWS, Google Cloud, or Microsoft Foundry. However, Anthropic acknowledged that signed metadata may not be available through every cloud platform or for every file type.
**Invisible watermarks follow copied Claude text**
When a supported model generates text, Claude inserts an imperceptible watermark into the output. Anthropic insists the mark does not change meaning, quality, or readability.
Because the watermark operates at the model level, it stays with the text regardless of which Claude product created it. The mark survives copy-and-paste actions and may endure light editing.
For files like SVG, PNG, and JPG, Claude instead applies digitally signed provenance metadata built on the C2PA open standard from the Coalition for Content Provenance and Authenticity. A valid label signals that Claude processed the file and may reveal whether it was altered afterward. Still, Anthropic stops short of claiming that label proves Claude originated everything within the file.
**A Claude mark signals processing, not authorship**
Claude often proofreads, translates, summarizes, or converts existing material. As a result, output may carry a Claude mark even when the core ideas, text, or data came from elsewhere.
Users can also edit, excerpt, or combine marked content with other material after it leaves Claude. A detected mark therefore indicates Claude may have processed the content, rather than establishing its complete history. The reverse holds true as well: failing to find a mark does not prove content was created without AI.
Heavy editing, paraphrasing, translation, or mixing with other writing can strip away the watermark. Very short passages may lack enough material for reliable detection, while provenance metadata can vanish when files get converted, resaved, or screenshotted. Content from older Claude models or unsupported platforms may show no detectable mark at all.
Anthropic is working to extend marking to models released before August 2 under the legislation’s transition provisions. The company is also developing detection tools for users and third parties, though it has not yet published technical documentation or a launch timeline. Developers deploying Claude within their own products remain responsible for determining how Article 50 applies to their services.
