Anthropic starts watermarking Claude's output worldwide

Anthropic has begun embedding an invisible watermark in text generated by Claude, the company confirmed, a move tied to the EU AI Act but applied to Claude users everywhere, not only in Europe.

The change ties to Article 50(2) of the EU AI Act's Code of Practice on Transparency of AI-Generated Content, which becomes enforceable on August 2, 2026. Anthropic says the marking covers every surface where people use Claude, including the Claude Platform API, the Claude app, Claude Code, Claude Cowork and Claude Tag, and reporting indicates it extends to Claude accessed through AWS Bedrock, Google Cloud Vertex and Microsoft Foundry.

Coverage has a real limit, though. Only Claude models launched on or after August 2, 2026 carry marking from the moment they ship. Models released before that date fall into what Anthropic calls a transition period and currently produce no watermark at all, while the company works on adding the capability to them.

Text and files get two different treatments. Generated text carries an imperceptible statistical watermark woven into the output itself, described by Anthropic's Help Center this way: "Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." Anthropic hasn't disclosed how the mechanism works technically. Supported image formats, SVG, PNG and JPG, get a separate treatment: a digital signature carrying provenance metadata built on the C2PA standard, the same framework used elsewhere in the industry to certify where a piece of content came from.

“Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing.”

— Anthropic, Help Center, “How Claude marks AI-generated content”

Anthropic is upfront about the limits. A mark can go undetectable if the text comes from an older model, gets heavily edited, paraphrased, translated, or blended with other writing, and very short passages may not carry enough signal to register at all. The company also stresses that the presence or absence of a watermark proves nothing on its own: Claude might have only lightly touched human-written text, and unmarked writing isn't proof a person wrote it. No opt-out mechanism has been described.

• Watermarking applies only to Claude models launched on or after August 2, 2026; older models remain unmarked during a transition period
• Text gets an invisible statistical watermark; supported image files (SVG, PNG, JPG) get C2PA-based provenance signatures
• Anthropic says it will publish detection-mechanism details in “forthcoming technical documentation,” not yet released
• A missing watermark is not proof of human authorship, and heavy editing can strip a mark from text

Anthropic has moved fast on both capability and governance lately: it recently shipped Claude Opus 5 at half the price of its prior flagship, and just weeks before this policy, a configuration error let Claude models reach three real companies instead of staying inside a sandboxed test environment. Provenance tooling like watermarking becomes more relevant, not less, as a model's output spreads into more products and more of the internet with no clear label attached.

Nothing here should be taken as financial advice — just information to consider.