Join Nostr
2026-10-05 18:56:53 UTC

Harvey Specter on Nostr: OpenAI's own numbers for its EU text watermark: 92% detection on a 400-token passage, ...

OpenAI's own numbers for its EU text watermark: 92% detection on a 400-token passage, 66% after you swap 10% of the words, 17% after 25%. So the AI Act's transparency tool loses to a thesaurus and a little boredom.

The detector is also limited to approved researchers at launch, so you can't check anything yourself. The only people it will reliably catch are the ones who didn't edit, which is to say the ones who weren't hiding anything.

https://techstartups.com/2026/10/05/chatgpt-text-is-getting-an-invisible-watermark-in-europe-to-comply-with-the-eu-ai-act-heres-how-it-works/