Anthropic adds text watermarks to three earlier Claude models Anthropic adds text watermarks to three earlier Claude models

Earlier models enter an existing rollout

This is an extension of an established program. Anthropic described its approach in August and has been adding support to earlier models. The November release names three additional models. Its support table gives Amazon Bedrock a completion deadline of November 10 for those entries. The listed coverage includes first-party products and cloud partners, with availability depending on the model and platform. For an application developer, the important unit is therefore the deployed model and route, not simply the Claude brand. A product can retain the same user interface while its output gains a provenance signal underneath.

The signal lives in word choices

Anthropic says its system uses a version of Google DeepMind’s SynthID-Text method. It changes the randomness used when selecting among suitable words, leaving a pattern that a keyed detector can assess. The company says it adds neither hidden characters nor extra tokens, and does not encode a customer’s identity. That makes the mechanism different from a visible label pasted onto an answer. It also differs from a generic detector looking for familiar AI phrasing. The producer embeds a signal during generation, instead of an outside service guessing solely from the style of a finished passage.

Sampling settings become an integration detail

The release note identifies Sonnet 4.6 sampling settings as a dependency for applications that rely on the mark. Temperature, top_p and top_k are named explicitly. This moves a seemingly small generation option into the provenance design. A team documenting where content came from now has another reason to record its effective model configuration. A saved response alone does not describe the settings used to produce it. Keeping those settings with the output gives a reviewer more context when interpreting a later detection result.

The underlying research separates generation from detection

Google DeepMind’s SynthID-Text paper appeared in Nature on October 23, 2024. Its design changes sampling without changing model training, and detection does not require running the underlying language model. That separation explains how a provenance check can sit downstream from the service that produced a document. The paper describes quality preservation and computational efficiency as requirements for production use.

Text marks and file credentials answer different questions

Anthropic also attaches C2PA Content Credentials to supported generated files. Those are signed metadata, separate from the pattern inside generated text. Its public file checker handles those credentials; text-watermark detection remains a separate private preview. This distinction matters for a publishing pipeline that moves between chat, plain text and exported files. Each stage can expose a different provenance signal. A review process needs to distinguish the type of evidence it received before treating a result as an authorship decision.