Imagine a writer who, every time he hesitates between two words with the same meaning, "start" or "beginning", "quick" or "swift", chooses according to a secret code. Nobody notices a thing. His sentences still read naturally. But someone who knows the code can pick out his writing from thousands of others.
That is, broadly, what OpenAI is going to do with ChatGPT's texts. And the first people affected are Europeans.
What has been announced
On Monday OpenAI presented a marking system called textGrain. It slips an invisible signal into the choice of words ChatGPT makes as it writes, using a secret key. The text stays exactly the same for the reader. But a detector can then establish, statistically, that it was produced by ChatGPT.
In the coming weeks, this marking will be applied automatically to texts produced by ChatGPT and by Codex, OpenAI's programming tool, for users in the European Union, whatever their plan. Elsewhere in the world, it will remain off by default: only developers who want it will be able to turn it on.
Why Europe first
Because the law requires it there. Since 2 August 2026, the European regulation on AI has obliged providers of generative AI to make their content machine-identifiable. We explained this in our article on the European transparency rules. Services already on the market had until 2 December to comply.
The United States has no equivalent rule at federal level. The result: a user in Paris and a user in New York will no longer have exactly the same ChatGPT.
What it changes for you
In practice, nothing visible. Your texts will look like they did before. And if you copy and paste a text from ChatGPT elsewhere, the mark travels with it, since it is written into the words themselves.
But not everyone will be able to read it. The detector is not public: OpenAI reserves it, case by case, for approved researchers and specialist bodies, under a European code of conduct. And according to the company, it only indicates whether the mark is present, without revealing who wrote the text, or the prompt, or the conversation.
In other words: your teacher or your boss will not be able, overnight, to run your texts through a detector to find out whether you used ChatGPT.
According to figures published by OpenAI, on a text of around 400 tokens, that is roughly three hundred words, the mark is detected in more than nine cases out of ten. But if one word in ten is replaced with a synonym, detection falls to around two cases out of three. With one word in four replaced, it drops to fewer than one case in five. And on a very short text, it becomes hard to spot. A simple rewrite is therefore often enough to make it disappear.
What the mark does not say
OpenAI admits it itself: a detected mark does not prove who wrote a text, nor that it is accurate, nor what a human changed in it. It is a clue, not proof.
It is the same lesson as with every tool of this kind. We saw it with AI text detectors, easily circumvented: recognising text written by a machine remains a problem nobody has fully solved.
What we take away
It is a fairly rare moment: a European law that concretely changes the product of the world's largest AI company, for Europeans only.
This watermark will not settle everything, far from it. But it shows that the rules written in Brussels do not always stay on paper. OpenAI has also said it wants to release its technology on an open-access basis, which would allow others to adopt it, and perhaps one day make it a standard well beyond Europe.