What Can a Claude Text Watermark Prove About Who Wrote Something?

Anthropic says future Claude models will leave an imperceptible statistical pattern in generated text. A detection API is coming later. Put a student essay, job application or edited article beside that announcement. A detector score could be treated as evidence against a person, so the caveats need to stay on screen. Anthropic says a match can only estimate whether Claude was involved. It cannot tell whether Claude drafted the piece or heavily edited someone’s work. Short passages give the detector less to work with. Proofreading, factual answers and code may carry little signal because there are fewer harmless word choices. A thorough rewrite can remove it. The mark also contains no identity, account or chat information. So a positive result should not become an automatic cheating or authorship verdict. Ask for the draft history, sources and the person’s account of how the work was made. A negative result is not proof that no AI was used either. If a teacher, editor or employer cannot handle both of those cases, the detector is being asked to make a decision the source says it cannot make.