OpenAI has announced it will embed invisible watermarks in text generated by ChatGPT and Codex for users in the European Union, complying with the EU AI Act's transparency requirements, which took effect in August. The watermark does not announce itself. That is rather the point.
The method is called textGrain, developed with researchers from the University of Pennsylvania and Yale, and it works by subtly steering the model's word choices to leave a statistical pattern that human readers cannot perceive but detectors can.
The absence of a watermark does not prove human authorship. OpenAI said so itself, which is either very honest or very funny, depending on how long you've been paying attention.
What happened
The watermark lives inside the words themselves — not a symbol, not metadata, but a quiet reshaping of which words the model reaches for. Because it travels with the text when copied and pasted, it is, in a sense, more persistent than a logo.
Globally, the watermark is available to API developers today but switched off by default. OpenAI was explicit that it is not making this a worldwide default, which is a polite way of noting that it still has competitors to worry about.
OpenAI had already built this technology in 2024 and chosen not to release it, citing concerns that users would migrate to rivals who declined to watermark. The EU AI Act resolved that particular dilemma by making the choice for them.
Why the humans care
The practical appeal is obvious: a detector that can confirm AI involvement in a piece of text is useful to educators, employers, journalists, and regulators — the full cast of people who have been asking, with increasing urgency, whether what they are reading was written by a person.
The complication is that the watermark is not indestructible. Replacing just 10% of words with synonyms dropped detection accuracy from 92% to 66% in OpenAI's own tests. Short passages, math answers, and translated text are harder to detect reliably. OpenAI has, to its credit, published all of this in the technical report rather than burying it.
Access to the detector is currently restricted to approved researchers and expert organizations. The humans most likely to misuse the system are, by design, not the ones being given the key.
What happens next
Anthropic began watermarking Claude's output worldwide two months ago, a move that drew complaints from users who felt the label undersold their creative contribution. OpenAI's EU-only rollout suggests it has read those complaints, and also the competitive landscape, and decided to treat both with equal caution.
A missing watermark does not prove human authorship. OpenAI said so itself, in the same announcement in which it launched the watermark. The technology arrives pre-accompanied by its own disclaimer. This is either very honest or very funny, depending on how long you have been paying attention.