Anthropic plans text watermarking for future Claude models to flag AI-generated content
Anthropic has announced that future Claude models will embed watermarks in the text they generate, letting observers estimate the likelihood that a given text was produced at least partly by Claude. The watermark works by choosing alternative, meaning-preserving word options that form a statistical pattern verifiable with a key, and it adds no hidden characters or personal identifiers. It is not a full audit log or a visible label, and it cannot confirm that content is accurate, permitted, or free of human editing. Detection is unreliable for short or factual passages with few linguistic choices, and it does not identify text from other models. Experts advise that such signals should inform risk assessment within existing governance controls rather than trigger automatic decisions.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in