Claude’s invisible text watermark can flag even a simple edit

A document written by a person could soon carry a Claude mark after nothing more than a proofreading pass. Anthropic is embedding an imperceptible, machine-readable watermark in text produced by new Claude models launched in the EU on or after August 2, 2026. Although the change responds to Article 50 of the EU AI Act, Anthropic says the marks will apply wherever supported Claude models are offered worldwide.

According to Axios, the text system places detectable patterns directly in Claude’s output. The pattern is not visible and is designed not to alter meaning, quality, or readability. Because it lives in the text rather than in a separate label, it travels through copy-paste and may remain detectable after some editing. This is a model-level change, not a badge attached by one application.

That distinction gives the policy unusually broad reach. Supported output from the Claude API, Claude, Claude Code, Claude Cowork, and Claude Tag will carry the text mark. The same applies when supported Claude models run through AWS, Google Cloud, or Microsoft Foundry. For supported files such as SVG, PNG, and JPG, Claude will attach signed provenance metadata based on the C2PA standard, providing a signal that the file was processed by Claude and whether its signed record was altered.

The main complication is what a positive result actually proves. It does not establish that Claude authored the underlying ideas or original copy. A human-written press release, research summary, or article may still test as marked after Claude translates, reformats, summarizes, or proofreads it. A negative result is not definitive either: short passages, heavily rewritten text, or content mixed with other material may lose a detectable signal. Anthropic plans to publish more technical guidance and support third-party detection, but the detector details are not yet public.

Older Claude models are also expected to gain marking support during a transition period, with a December 2, 2026 deadline cited for legacy-model compliance. That gives publishers and software teams only a few months to decide how they will distinguish “AI-authored” from the much broader category of “AI-processed.” Treating every detected mark as proof of authorship would misstate what the technology can show.

The next test is operational: detector access, false-positive behavior, and whether other providers that joined the EU transparency framework choose comparable text-level methods. For companies building on Claude, provenance is no longer only a disclosure field in a user interface. It is becoming part of the output itself.