Anthropic says future Claude models will mark generated text using a version of Google DeepMind’s SynthID-Text method. It changes how the model chooses among plausible next words, leaving a statistical pattern rather than adding hidden characters.
A detector with the matching key can estimate whether Claude was involved in a passage. It cannot establish that text was human-written or distinguish a fully generated document from one Claude edited heavily. Anthropic says the pattern contains no user, organization or conversation identifier.
The company reports no practical effect on readability or creativity in its internal tests. It also says the process adds no extra tokens.
Detection is weaker with short samples, light proofreading and code that leaves little room for alternative tokens. Substantial rewriting can remove the pattern. These limits make the watermark evidence of possible Claude involvement, not proof of authorship.