Claude Watermark Critique Loses on Prose, Wins on Privacy

Gruber called Claude's SynthID watermark a "perversion of writing" in a Daring Fireball post that drew 368 points and 348 comments on Hacker News.

Most of the thread didn't buy the prose complaint. Commenters pointed out that the watermark only biases word choice at high-entropy points, where several next words are close to equally likely. Google's own A/B tests reportedly found no difference in reader preference between marked and unmarked text.

The objection that survived was different. Checking whether a document carries the watermark means uploading the whole file to Anthropic. For anyone testing a draft of unpublished research or an internal memo, that's a privacy question, not a writing-quality one.

Anthropic's August 15 post added the specifics that shaped the debate. Light editing won't strip the mark, but a full rewrite where every word gets replaced will. A detection API is planned. Code stays largely unaffected, aside from what shows up in comments.

Prose isn't the real fight. Once a detection API ships, the question stops being whether a paragraph reads like Claude. It becomes who has to hand Anthropic a document to find out.

Anthropic's detail post landed August 15, two days before the Hacker News debate peaked at 368 points.