Anthropic Starts Watermarking Claude's Text Output, Under EU Pressure

Anthropic has started embedding invisible watermarks into text generated by new Claude models, effective August 2, 2026. The change is tied to Article 50 of the EU AI Act, which requires disclosure of AI-generated content, and Anthropic has signed the EU's Code of Practice covering that requirement — but the company says the marking applies to Claude output worldwide, not just to EU users.

The watermark is embedded at the model level, so it isn't something a specific app or interface adds on top. That means it reaches Claude output wherever a new Claude model is doing the generating: the API, Claude's own apps, Claude Code, Claude Cowork, and Claude Tag, plus Claude models accessed through AWS, Google Cloud, and Microsoft Foundry. The mark is invisible during normal reading, and Anthropic says it's designed to survive being copied and pasted, with some marks expected to survive light editing too.

Files get a different treatment than text. Supported file formats, including common image types, can carry signed provenance metadata based on C2PA (the Coalition for Content Provenance and Authenticity) — an existing industry standard for recording how a piece of content was produced. Anthropic is treating text and files as separate problems on purpose: a text watermark travels with copied content in a way file metadata doesn't, while file metadata can be stripped by something as ordinary as a screenshot or a format conversion, which a text watermark generally survives.

Older Claude models aren't getting this retroactively. Anthropic says it's working on adding marking capability to them during the EU AI Act's transition period, without giving a specific timeline in what's public so far.

Anthropic is also explicit that neither signal proves authorship on its own, in either direction. Claude can process text a person wrote — translating, summarizing, or editing it — and the output can still carry a Claude mark despite a human having done most of the original thinking. The reverse holds too: unmarked text isn't proof a human wrote it, since heavy editing can degrade a watermark and short passages may not carry enough signal to detect reliably. Anthropic says it plans to publish technical detection details so users and third parties can check content for supported marks themselves.

Why this is worth watching

This is my read, not something the source article claims directly: the EU AI Act keeps functioning as the de facto global floor for AI transparency rules, the same way GDPR became the default privacy baseline well outside the EU. Anthropic didn't need to make this worldwide — nothing in Article 50 requires marking content served to a non-EU user — but building one watermarking system for every user is operationally simpler than maintaining a geofenced version, and it reads better publicly than "we only prove our own text is AI-generated for European regulators." Whether the watermark actually holds up against determined removal, or whether it just becomes another easily-stripped layer like most digital watermarking has historically been, is the open question this story doesn't yet answer.

Source: Interesting Engineering