What OpenAI announced
The announcement is a response to Article 50 of the EU AI Act, which since 2 August 2026 has required providers of generative AI systems to mark text, images, audio and video “in a machine-readable format and detectable as artificially generated or manipulated”. OpenAI already marks images and audio. Text was the gap, and the company's post describes a three-part, phased answer.
First, within weeks, eligible text from ChatGPT and Codex in the EU will be watermarked for users on every plan. OpenAI does not define “eligible” in the post. Second, API customers anywhere can switch watermarking on for selected models from 5 October; it is off unless they do. Third, researchers and expert organisations can apply to use the detector. OpenAI says it will reconsider each part “as the technology, standards, and evidence evolve”.
The European Commission's Code of Practice on transparency of AI-generated content, which the Commission and the AI Board have accepted as a way to show compliance, sets out how providers can meet the marking duty. The Commission says about 190 organisations had signed it by the end of July 2026. OpenAI's post cites the code as the basis for limiting detector access case by case. Whether OpenAI has signed it is not stated in the post.
How the watermark works
A language model writes one token, a word or part of a word, at a time, choosing from a set of likely candidates. textGrain ties those choices to pseudorandom values generated from a secret key and the preceding text. The result reads normally but carries a statistical pattern. A detector holding the key can test a passage for that pattern; without the key, nobody can. The technical report, written by OpenAI staff with academics from the University of Pennsylvania and Yale, says the detector needs only the text and the key.
The strength of the signal depends on how much freedom the model had. Where many words would do, as in an explanation of a psychology concept, there is room to embed a pattern. Where the wording is constrained, as in mathematics, there is less. Length matters for the same reason: more tokens, more evidence.
OpenAI reports no meaningful loss of quality: across eight benchmarks for its Astra model, scores with and without watermarking were within a few points of each other, some higher and some lower. It also says textGrain matched or beat other methods it tested, including Google DeepMind's SynthID for text, while warning that strong results in ideal conditions do not guarantee reliable detection in everyday use.
What it can and cannot detect
OpenAI's own figures are the clearest account of the limits. At a target false-positive rate of 1 per cent, meaning about one human-written passage in a hundred wrongly flagged, the detector found the watermark in about 80 per cent of 200-token passages and about 95 per cent of 400-token passages on psychology questions. On mathematics the rates were substantially lower.
Share of watermarked passages detected at a 1% target false-positive rate, OpenAI evaluations
Source: OpenAI, Our approach to EU text provenance rules, 5 October 2026, read in full through a text-rendering proxy 5 October 2026.
Figures are OpenAI's and have not been independently verified. The first two rows and the last three come from separate evaluations. OpenAI says detection was substantially lower for mathematics, where word choice is constrained.
Editing is the bigger weakness. Swapping one word in ten for a synonym brought detection on 400-token passages down from about 92 to 66 per cent. Swapping one in four brought it to 17 per cent. OpenAI also says translation and very short passages may defeat detection, and that text from other companies' tools, from older or unsupported models, or written before watermarking began will not carry the signal.
The company is explicit about what a positive result does not mean. A watermark can indicate that an OpenAI system generated or processed part of a passage. It does not measure how much human judgement went into it, establish ownership or responsibility, identify the account or prompt, or say anything about whether the text is true. A negative result does not prove a human wrote it.
What this means for a company that drafts with ChatGPT
Who can check is narrow. For now only approved researchers and expert organisations can run the detector. An editor, a reader or a competitor cannot test a press release or a bylined article for an OpenAI watermark, and neither can the company that published it.
Where it applies is narrow too. The ChatGPT watermark is for the EU only at launch. A team drafting in ChatGPT outside the EU is not affected unless OpenAI widens the rollout; a team building on the API decides for itself.
Ordinary editing erodes it. A draft that is rewritten, shortened or merged with other material is the case OpenAI's tests show the detector handles worst. That is not an argument for disguising AI text; it is a reason not to treat a watermark, present or absent, as proof either way.
Disclosure is a separate duty. The AI Act's watermarking rule falls on providers such as OpenAI. The duty to disclose AI-generated text published on matters of public interest falls on the publisher and lapses where a person has reviewed it and holds editorial responsibility, as set out in our guide to the AI Act's labelling rules. Publications' own policies on AI-assisted copy matter more to a placement than any watermark, and Google's own guidance now asks sites to fact-check AI content before publishing it, as covered in what Google changed in its content guidance.
Methodology and limitations
This explainer is built on OpenAI's post “Our approach to EU text provenance rules” of 5 October 2026, which refused our direct requests and was read in full through a text-rendering proxy, and on the accompanying technical report “textGrain: Entropy-Calibrated Watermarking for Language Model Text”, dated 5 October 2026, read directly. The detection and editing figures are OpenAI's own evaluations, which we cannot verify; OpenAI says its charts use responses to questions from the ELI5 dataset. The chart shows only the figures OpenAI gives as numbers in its text, not values read off its charts. Article 50 and the Code of Practice are described from the European Commission's page on the code, read 5 October 2026, and from our earlier guide. We did not test ChatGPT or the API, have no access to the detector, and do not know which outputs OpenAI will treat as eligible.
Conclusion
Text from ChatGPT will start carrying a hidden mark in the EU within weeks, because the AI Act now requires machine-readable marking and OpenAI has chosen a statistical watermark to meet it. For a company that drafts with ChatGPT, little changes in practice for now: the mark is regional, the detector is in researchers' hands only, and OpenAI's own tests show that ordinary editing removes most of the signal. The more durable point is OpenAI's own list of what a watermark cannot tell anyone. It cannot say who wrote a passage, how much a person changed it, or whether it is true, and those are the questions an editor actually asks.
Frequently Asked Questions
Will ChatGPT watermark my text?
If you use ChatGPT in the European Union, OpenAI says eligible output will carry an invisible watermark within weeks, on all plans. Outside the EU it is not the default; API customers anywhere can opt in.
Can someone check whether text came from ChatGPT?
Not publicly. OpenAI is giving detector access only to approved researchers and expert organisations, initially case by case, because of the risk of false positives and missed watermarks.
How reliable is OpenAI's text watermark?
In OpenAI's tests at a 1 per cent false-positive target, it detected about 80 per cent of 200-token and about 95 per cent of 400-token passages on psychology questions, and substantially fewer on mathematics.
Does editing remove the watermark?
Largely. Replacing 10 per cent of words with synonyms cut detection on 400-token passages from about 92 per cent to 66 per cent, and replacing 25 per cent cut it to 17 per cent, in OpenAI's own evaluation.
Does a watermark show who wrote the text?
No. OpenAI says a watermark does not identify the user, measure human contribution, establish ownership or responsibility, or verify accuracy, and that the absence of one does not prove a human wrote the text.
Why is OpenAI doing this?
Article 50 of the EU AI Act, applicable since 2 August 2026, requires providers of generative AI systems to mark output in a machine-readable way so it can be detected as AI-generated.
Sources
- official OpenAI, Our approach to EU text provenance rules, 5 October 2026. Refused our direct requests; read in full through a text-rendering proxy on 5 October 2026.
- documentation Xiang Li, Garrett Wen, Xiaohong Chen, Qi Long, Arzav Jain, Florent Joly, Mike Lam, Qingquan Song and Weijie Su, textGrain: Entropy-Calibrated Watermarking for Language Model Text, OpenAI technical report, 5 October 2026. Read 5 October 2026.
- regulator European Commission, Code of Practice on Transparency of AI-generated Content. Article 50 obligations from 2 August 2026; about 190 signatories by end of July 2026. Read 5 October 2026.
- our research Media Morpheus Research, Does the EU AI Act Require You to Label AI-Written Coverage?, 25 September 2026.