Anthropic details text watermarking for three Claude models: what detection can establish

An AI repeatedly chooses what comes next in a sentence. When several expressions are similarly appropriate, there is room to choose. Anthropic describes using a rule for those choices, then checking the accumulated pattern across a passage. How the watermark works

“The meeting takes place this afternoon” and “This afternoon, the meeting takes place” can convey the same schedule. These invented sentences illustrate room for choice; they are not an actual watermark rule or detection sample. The developer says its mark adds no hidden characters or extra tokens and contains no identifying information about the user, organization or conversation. Text watermark documentation

%% The role of a text mark · simplified from Anthropic's explanation
flowchart TD
  A["Claude: choose the next expression<br/>Where suitable alternatives exist"] --> B["Choice rule accumulates in the text<br/>An imperceptible statistical pattern"]
  B --> C["Authorized detection tool<br/>Use a key to check the pattern's signal"]
  C --> D["Clue: possible Claude involvement<br/>Authorship and accuracy need other evidence"]

The text watermark is part of word selection. Files can instead carry Content Credentials following C2PA, an open standard for signed provenance metadata. Metadata is descriptive information stored with a file; a credential can record processing involvement.

A Claude-issued credential on a supported file such as PNG or JPEG supplies a signal that Claude processed it. Copying text and re-saving a file can affect different kinds of evidence, so first identify which mark a tool checked.

The name Claude covers different models and access routes. Coverage depends on the exact model and whether it is used through a Claude app, the API or a cloud partner. An API lets another program send requests to the model and receive responses.

The Sonnet 4.6 exception names temperature, top_p and top_k, settings that adjust generation choices. Developers relying on the watermark are told to omit all three, including values inserted by a framework or client library.

Consider a fictional newsletter whose author asks Claude to proofread or translate a human-written draft. A detected mark can indicate involvement in processing. It does not reconstruct who first supplied the ideas, research or every sentence.

Light proofreading may leave a weak signal because Claude selected few words; a complete translation can involve many more selections. Even then, detection does not determine factual accuracy, authorship and ownership together. A recipient can compare the original draft, revision history and cited material to investigate those questions more concretely.