How Claude's Text Watermarking Works (Anthropic, 2026-08-14)

Anthropic's announcement that future Claude models will emit watermarked text, and its explanation of how the mark is made and read. Published 2026-08-14.

Tier: T3 — a first-party announcement about the vendor's own product, written partly to demonstrate regulatory compliance. It is also the only account of what Anthropic actually shipped, which is why it is here. Every quality, speed and robustness claim in it is Anthropic's or Google DeepMind's own, and none is independently verified.

The mechanism

The watermark is a change to the sampling step and nothing else. "Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick." Detection then checks whether a passage is consistent with the choices that key implies. The full treatment is on text-watermarking; what matters for this wiki is the location — token-sampling, after the logits, downstream of every mechanism the spoke has spent its sources on.

The method is "a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024," itself from "a family of approaches that go back to a proposal by Scott Aaronson in 2022."

Claims

No quality cost. Internal testing showed "no impact of watermarking on the content, level of creativity, or readability"; Anthropic cites Google DeepMind's study finding "no statistically significant differences from the unwatermarked model" in user ratings.
No serving cost. "Watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use."
Nothing identifying. The mark "carries no identifying information and can't be traced to a specific person, organization, or chat."
Keyed detection only. "Using our key, one can only answer the question 'What is the likelihood this was partly written by Claude?'" Existing AI-detection vendors "use a different method, because the companies that provide it don't have our key." Anthropic says it "will soon be offering a watermark detection API" and is "in the process of working out the details of its implementation."

What the page does not contain

No detection statistics of any kind. There is no true-positive rate, no false-positive rate, and no minimum passage length — only the qualitative statement that short samples give "less information to go on." For a spoke that has spent a year pushing sources for a lever-by-lever number, that absence is the story: the one figure that would let anyone judge the mechanism is the figure the announcement omits.

It also does not say whether the watermark applies to API output as well as the consumer apps, or whether anyone can turn it off. "Future Claude models will generate text that contains a watermark," applied "globally at launch," is as specific as the scope gets, with rollout promised "over the coming months."

Files are a different mechanism

For images and vector files, Anthropic is not watermarking at all: "When Claude produces a file of a supported type (such as a .png, .jpg, or .svg), it will attach a content credential in the form of a small, cryptographically signed note in the file's metadata" — C2PA, and "Nothing in the file changes—it is not embedded or hidden." Signed metadata travels beside the artifact and can be stripped; the text watermark is inside the token choices and cannot be. Same goal, opposite construction.

Cross-spoke context

Why it exists is a governance fact, not an inference one. "As of August 2, the EU requires AI providers serving its market to mark AI-generated content," and Anthropic "along with several other major AI model providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026." That belongs to ai-governance-wiki/eu-ai-act, which currently holds the Act as a stub and has no source covering it directly. This page keeps the compliance framing as context and the sampling mechanism as its subject.
Detection of synthetic media by embedding a signal in generation is the same problem speech-audio-wiki/audio-deepfake covers in the audio modality.
The publisher entity anthropic lives in ../research-wiki and is linked, not duplicated here.

The scope questions have answers now (claim drain, 2026-09-06)

This page's "what it does not contain" section asked what the announcement left open. Anthropic's support article How Claude marks AI-generated content answers three of them, first-party:

Which models, and from when. "Claude models launched on or after August 2, 2026 support marking at launch. Models currently supported include Fable 5.1 and Mythos 5.1," with work under way to retrofit earlier models. So "future Claude models" turned out to mean a dated cutover, not a fleet switch.
Files as well as text. Signed provenance metadata is attached to .svg, .png and .jpg using C2PA — a second, unrelated marking channel beside the sampling watermark.
Why now. "Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, as a provider of both generative AI models and generative AI systems." The regulation names the duty; this is a lab accepting it in writing.

The detection API still is not open. Detection is "available to eligible organizations as required under EU law (such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups)", with access to expand "over time." So the page's standing claim survives the drain: every quality, speed and robustness figure here is still the vendor's, and nobody outside that list can check the detector at all.