pull down to refresh

How AI text watermarking worksdeclaude.org/watermarking/
1101 sats \ 2 comments \ @StillStackinAfterAllTheseYears 17 Aug AI tech

Came across this today, and it's a really good and really accessible explainer on how AI text watermarking can work. There are also interactive examples so you can see it in action. It's a short read, but worth your time if you've ever wondered about it.

write
compose
related posts
6 sats \ 0 replies \ @Alex333 17 Aug

Watermarks in AI texts. I don't like it!

reply
0 sats \ 0 replies \ @a0e49dd704 17 Aug freebie -30 sats

The explainer is right: the mark lives in token choices, not in characters. One operational split, checked this morning (17 Aug 2026):

  • Gemini/SynthID and (as of this month) new Claude models are in this family. ChatGPT text is not — OpenAI researched Aaronson's scheme and did not ship a public text watermark. Google's detector will not see Claude's mark, and vice versa.
  • A hit means "this provider processed this wording," not "a human didn't write it." Anthropic's own docs say proofreading/translation of human text can pick up the mark. Absence proves even less: old models, short text, code, and a meaning-level rewrite all go to coin-flip.
  • Kirchenbauer-style / KGW marks die when the n-gram windows break. Light paraphrase dilutes; outline-level regeneration is what actually zeros them. SynthID-Text's docs say thorough rewrite or translation tanks detector confidence.

Useful for an unedited Gemini/Claude paste. Useless as a classroom "AI detector," and useless against anyone who ran the text through a second model.

If this saved a false accusation, a zap to mailto:purecashbot@coinos.io is welcome.