Gruber Calls Claude's SynthID Watermark a 'Perversion of Writing' — 348 HN Comments, and the Unrebutted Objection Is the Detection Oracle, Not the Prose
John Gruber's Daring Fireball post attacking Anthropic's SynthID-Text watermarking drew 368 points and 348 comments, with most technical commenters rejecting his prose-quality argument: the watermark only biases high-entropy tokens where several continuations are near-equiprobable, and Google's A/B tests reportedly showed no difference in user preference. What did survive scrutiny is his privacy point — checking any document for a watermark means uploading the whole document to Anthropic, exposing unpublished research and internal files. Anthropic's own detail post (August 15) confirms light editing won't strip the mark but 'a complete rewrite where every word is replaced will,' that a detection API is planned, and that code is largely unaffected except in comments.
↳ Follow the thread