ChatGPT text in the EU is getting an invisible watermark. It does not do your labelling for you.

On 5 October OpenAI said it will add an invisible watermark to text that ChatGPT and Codex write for users in the European Union. The rollout runs over the coming weeks and covers all plans, from free accounts to Enterprise. Outside the EU nothing changes by default. OpenAI gives one reason: the AI Act requires providers of generative AI to make generated text identifiable by machine, and those transparency duties have applied since 2 August 2026.

The technology is called textGrain. It nudges the random choices the model makes between possible next words, so that a long passage follows a pattern a detector with the right key can spot. OpenAI says it adds no hidden characters, invisible spaces or odd punctuation. A reader sees normal text. On OpenAI’s own benchmarks the watermark made no meaningful difference to answer quality.

What the mark does not say

OpenAI is unusually direct about the limits, and the limits are the useful part for an employer. The watermark carries no information about the user, the organisation, the account or the prompt. It cannot tell how much a person edited the text. It says nothing about who owns the text or whether a disclosure was required. And a missing watermark proves nothing, because short, edited or translated text often loses the signal.

The published numbers show how quickly that happens. At a 1% false positive rate, the detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages on a subject like psychology, and far less often in mathematics. Swapping 10% of the words for synonyms cut detection from about 92% to 66%. Swapping a quarter of them cut it to 17%. OpenAI ran its editing tests on English text and has published no figures for Estonian.

For those reasons the detector is not public. Approved researchers and expert organisations can apply. OpenAI’s help page states that general business uses, such as fraud checks, sit outside the programme, and so does checking student work for AI use. API customers anywhere can now switch watermarking on for selected models in their project or organisation settings. It stays off by default.

Where this sits in the AI Act

The duty to mark output belongs to the provider, which here is OpenAI. The European Commission published the final Code of Practice on transparency of AI-generated content on 10 June 2026, and OpenAI says it signed it. Every other provider of generative AI serving EU users faces the same marking requirement under Article 50. The deployer duties we covered in September did not move: if your company publishes AI-written text to inform the public on a matter of public interest, you label it unless a person reviewed it and someone carries editorial responsibility. A watermark inside the words does not count as that label.

What an Estonian company should do this week

• ✳Tell staff plainly what changed. Their ChatGPT text will carry a mark that does not name them or the company, so there is nothing to hide and no reason to stop using approved tools
• ✳Keep your disclosure routine as it is. Where you label AI-written public content today, keep labelling, and keep the record of who reviewed it. The watermark changes neither
• ✳Write one rule into your AI policy: no decision about an employee, candidate, supplier or student rests on an AI detector result. OpenAI’s own detector is not open to business use, and its maker says a negative result proves nothing
• ✳If you build a product on the OpenAI API that serves EU users, open the project settings, find the watermark option and decide with whoever owns your AI Act compliance whether to switch it on
• ✳Ask your other AI vendors, Microsoft, Google and Mistral among them, how they mark text for EU users. The requirement is the same for all of them, and their answers will differ

The short version for Monday: OpenAI has handled its part of Article 50 for text. Your part, labelling what you publish and treating detectors with suspicion, is still yours to do.