- Sources: primary, discussion
- Summary: The method biases only low-stakes token choices by seeding the sampling randomness from a key and the preceding words, and Anthropic states there are no hidden characters, no extra tokens and no price change. Code carries less watermarking than prose because exact output leaves no free choice, and factual passages are sparser for the same reason. Anthropic states that light editing probably will not remove the mark, that a complete rewrite will, and that the detection API has not yet shipped.
- Why it matters: This answers the mechanism question this archive attached as a follow-up to its 2026-08-11 entry, and sets the practical limit that any rewriting stage in a pipeline erases the mark.
send feedback on this story