Claude Now Whispers an Invisible Signature Into Every Word It Writes

Since August 2, 2026, every reply from Claude carries a hidden mark. Not a banner, not a disclaimer. A signature woven into the text itself.

Anthropic now injects an invisible watermark into all responses from supported models. The change aligns with Article 50 of the EU AI Act, which requires providers of generative AI to make synthetic content machine-identifiable. Anthropic signed the associated code of practice and applied the mechanism globally, not just in Europe.

How a watermark hides in plain sight

For images, a digital watermark is straightforward. For text, it is harder. Claude does not insert hidden characters or strange spacing. Instead, it biases word selection ever so slightly. The result is a statistical pattern that an automated tool can detect on a sufficiently long passage, but which leaves no trace a human would notice.

The watermark isn’t added to the text; it is the text, hidden in how the words were chosen.

Think of it as an instruction to use a certain letter every 25 characters. The pattern is invisible to the eye. A machine, however, can count the occurrences and spot the signature. Anthropic’s real method is more sophisticated, but the principle is the same.

Where the watermark fails

A statistical fingerprint does not survive a rewrite. Ask another model to paraphrase a Claude output and the signature dissolves. The system catches only the user who does not know, or does not care, to mask the origin. A few minutes with any other language model is enough to scrub the mark.

False positives are also a risk. A human who accidentally writes with a similar letter distribution could match the pattern. And the watermark cannot hold through translation or deep editing.

A silent surveillance layer

The most uncomfortable detail is the absence of notification. Users composing private correspondence, drafting a report, or writing an email with Claude are not told the output bears a traceable mark. Their words become identifiable as AI-generated without their knowledge. As Numerama points out, this silently inserts a surveillance layer into personal communication.

Anthropic’s move is logical under the new regulation. It is also a quiet reminder that invisible marks cut both ways: they help platforms comply, but they also bury a signal in every sentence you send.