A watermark that one word in ten defeats, and a detector almost nobody has

OpenAI started embedding an invisible mark in ChatGPT and Codex output on 5 October, 64 days after the EU AI Act's transparency obligations began applying on 2 August. The system is called textGrain, it covers all plans, and it runs in the EU only. Developers elsewhere can switch it on through the API, where it ships off by default.

“initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses”

— OpenAI, Заявление 5 октября 2026

OpenAI published the number that limits it

The mark works by «subtly shaping the model's word choices, leaving a pattern readers can't see, but a detector can pick up». It survives copy and paste, it does not identify the user, and OpenAI measured no meaningful change in model performance.

The OpenAI text watermark lives inside the LLM's choice of each next word, which is why it survives copy and paste and why a synonym pass damages it. Nothing is appended to the file and nothing sits in the metadata, so there is no header a reader could inspect.

Detection runs at about 92% on untouched text. Replace 10% of the words with synonyms and it falls to 66%. That is 28% of the detection rate gone for a light paraphrase. Short passages, maths answers and translated text are harder again, and OpenAI states that a missing watermark does not prove human authorship.

Who can read the mark

Only approved researchers and expert organisations get the detector. A teacher reading an essay cannot run it. Neither can an editor or a court.

The obligation puts a signal in the text and leaves the people who care about the text unable to read it. OpenAI is candid about why, framing the restriction as a way to evaluate reliability and responsible uses first.

Geography decides which product you get

The same model from the same company behaves differently depending on where the user sits. In the EU the mark is on and nobody opts in. Everywhere else it is a switch that ships off.

We covered the mirror image on 24 September, when the iPhone 18 Pro shipped to 65 countries with its AI features switched off in the EU and China. The usual pattern gives Europe less. This one gives Europe a disclosure the rest of the world does not get by default, and the thing being disclosed is the same text either way.

textGrain marks machine writing, loses about a quarter of its accuracy to a thesaurus, and can be checked only by people on OpenAI's list. It joins the other things a model does to your text that you cannot see, which we wrote about yesterday.

Informational material, not investment advice. The detection figures are OpenAI's own published numbers, and no independent evaluation of textGrain exists yet.