Anthropic's 'watermark' text adulteration in Claude is a perversion of writing
Points and comments are a snapshot, not live.
Anthropic's text watermarking corrupts word choice, violating promises of imperceptibility.
Anthropic's Claude will watermark all text over 200 tokens by biasing token selection toward a secret 'green' list, making statistically detectable word choices. The author argues this adulterates meaning, as no synonyms are exact substitutes. Anthropic's original support document claimed no impact on quality, but the actual steganographic technique changes word selection. Google's SynthID similarly adjusts probability scores, described as unnoticeable despite clear semantic differences like 'bananas' vs 'pineapple'. The EU regulation motivating this is called 'red-tape nanny-state pipe-dream nonsense' that will only catch honest users while evaders use paraphrasing tools like Declaude.
What commenters are saying
Commenters split: some defend watermarking as reasonable for detection, others call it security theater. A top comment notes the technique only applies to prose, not code or structured output, citing Anthropic's own documentation. Several point out that RLHF already alters word choice, questioning the purity objection. A correction: proofreading that flags errors without rewriting avoids false flagging, but copy-pasting AI edits would be detectable. One commenter argues LLM outputs should be flagged because they are uncopyrightable public domain, while human works retain copyright.