Anthropic starts invisibly watermarking Claude's text and image outputs

Anthropic said Claude models launched on or after 2 August 2026 embed an imperceptible, machine-readable watermark directly into generated text, alongside signed C2PA provenance metadata attached to generated files such as images. The company said the change applies worldwide across the API, Claude Code and cloud partners, and framed it as fulfilling commitments under Article 50 of the EU AI Act’s Code of Practice on transparency, which it had signed.

In a post setting out the method, Anthropic said the text watermark works at the level of word choice: wherever several wordings would carry the same meaning, the model biases which one it picks using a value derived deterministically from a secret key and the preceding text, rather than choosing at random. The pattern is invisible to a reader — “nothing is added to the text and there are no hidden characters,” the company said — but is recoverable by anyone holding the key. The technique is based on Google DeepMind’s SynthID-Text method, published in 2024, and applies by default to text from Claude models released from 2 August onward. Because the signal rides on genuine choices between interchangeable words, Anthropic said it is sparse where only one answer fits — a factual statement, or code — and unreliable on very short passages; it also said the marking does not affect output quality.

Anthropic’s support documentation describes the text watermark as designed to travel with the content when copied and pasted elsewhere, and to survive light editing, but said it would degrade if text is heavily rewritten, paraphrased, translated or mixed into other writing. File-level C2PA metadata is similarly fragile, and can be lost through format conversion, re-saving or screenshots. Models released before 2 August do not carry the feature yet; Anthropic said it was working to retrofit it. Detection tooling for third parties to check content for the marks was described as forthcoming, with technical documentation to follow.

The move follows OpenAI’s adoption in May of Google DeepMind’s SynthID watermark for images paired with the same C2PA standard, making watermarking-plus-provenance-metadata the emerging norm among major labs facing the same EU transparency deadline. As with OpenAI’s rollout, Anthropic cautioned that the signals indicate Claude’s involvement in producing content rather than serving as proof of full authorship, and that a determined user can still strip both marks.

Independent outlets first reported the update on 11 August, drawing on Anthropic’s undated support page; Anthropic set out the method more fully in a dated post on 14 August. A public detection API, which the company said would let third parties check whether text carries the mark, was still described as forthcoming.