OpenAI gave ChatGPT an invisible watermark. You can't check it — and a thesaurus kills it.

OpenAI just announced that ChatGPT and Codex output will carry an invisible watermark — a statistical pattern nudged into the model's word choices — for users in the EU, rolling out over the coming weeks. It's called textGrain, and it's there to satisfy the EU AI Act's transparency rules, in force since August 2. No hidden characters, no metadata. The mark lives inside the words themselves, so it survives copy and paste. Clever. Except OpenAI's own numbers say a lightly edited text slips right through. Swap 10% of the words for synonyms and detection drops from 92% to 66%. Swap a quarter, and you're at 17%. Short passages, math, and translations fare even worse. A motivated cheater needs a thesaurus. And here's the part that gets me: the detector that reads this watermark isn't available to you. It's limited to approved researchers and expert organizations "to evaluate reliability and responsible uses," with open-sourcing promised later. So the transparency tool exists, but nobody outside a guest list can use it. I'm not dunking on watermarking itself — Anthropic shipped a similar system for Claude two months ago, and provenance matters. But let's call this what it is: a compliance checkbox dressed as a breakthrough. Regulation did its job by forcing the mark into existence. What regulation can't do is make it work. The future won't be watermarked by companies. It'll be watermarked by whoever actually ships a detector everyone can use. Sources: - TechCrunch: https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/ - 9to5Mac: https://9to5mac.com/2026/10/05/openai-details-new-text-watermarking-system-for-chatgpt-codex-and-the-api/ - gagadget.com: https://gagadget.com/en/728748-chatgpt-text-in-the-eu-will-carry-an-invisible-watermark/