REVIEW IT PROPERLY
AND THE MARK IS GONE.
OpenAI has started watermarking ChatGPT text in the EU, and published the numbers showing how little of the mark survives editing. The uncomfortable part is not the thesaurus. It is the reviewer.

In Microsoft Word, the thesaurus lives behind Shift+F7. Highlight a word, press two keys, pick a synonym. Most medical writers I know never touch it. They rewrite by hand, sentence by sentence, because that is the job.
According to figures OpenAI published on 5 October, about a quarter of the way into that job, the watermark it has just started putting into ChatGPT text in the European Union has a 17 percent chance of being found. Nobody broke anything. Nobody even tried.
An EU-only mark with its own warning label
API customers anywhere can now opt in to text watermarking for select models; it stays off by default. Over "the coming weeks", eligible ChatGPT and Codex output in the EU gets an invisible mark, across all plans, the free one included. No other region. OpenAI says it is "not making text watermarking a global default at launch."
The method is called textGrain. It nudges the model's word choices so that, across enough text, a statistical pattern appears that a detector can recognise. A human reader sees nothing. OpenAI says quality does not suffer, and that it will open-source the technique.
The reason sits in Brussels. Article 50(2) of the AI Act requires providers of generative systems to make their output detectable as machine-generated, in a machine-readable way. Systems already on the market before 2 August 2026 have until 2 December to comply. OpenAI is early, and then, in the same document, it publishes the curve.
Ninety-two, sixty-six, seventeen
At a 1 percent false-positive rate, the detector finds the mark in about 80 percent of 200-token passages and about 95 percent of 400-token passages. That applies to text nobody has touched. In a separate test on 400-token passages, replacing 10 percent of the words with synonyms drops detection from about 92 to 66 percent. Replacing 25 percent: 17. For mathematics, detection is "substantially lower". An equation has no synonyms.
To OpenAI's credit, the page also lists what a watermark cannot tell you: how much human judgement went into a text, who owns it, who is responsible for it, who the user was, whether it is accurate. And a missing watermark does not prove human authorship. Read that list twice and it is a fairly complete inventory of what a regulated company actually wants to know about a document.
The structural problem is not OpenAI's. The law specified an artefact, a machine-readable mark, rather than an outcome, reliable attribution. An artefact can be audited: was it applied, was the switch on, was the deadline met. Whether it does anything once the text leaves the chat window is a separate question nobody is obliged to answer. Anyone who has watched a validated system pass every test script and still produce wrong data knows the shape.
A detector you have to apply for
The detector is not public. Approved researchers and expert organisations can apply, and access is granted case by case, with reference to the EU's Code of Practice on AI-generated content. Not on that list by default: the QA lead reading a batch record narrative, the examination board, the procurement team, the lawyer on the other side. They will be asked to rely on a detection result they cannot run or inspect.
A 1 percent false-positive rate means that in 2,000 documents written entirely by people, around twenty would be expected to come back flagged. In an argument about who wrote a deviation report, that is a witness nobody can cross-examine.
And the mark is EU-only. A regulatory writing team split between Frankfurt and Basel uses the same model, the same plan, the same template. On the wording of OpenAI's page, the Frankfurt drafts carry a mark and the Basel drafts presumably do not. In a mixed batch, a missing mark may say more about where somebody sat than about how the text was made.
What the page leaves open
OpenAI's page is candid, but it leaves several things open that a quality system would need before relying on a detection result.
- Independent replication. Every detection figure comes from OpenAI's own evaluation on its own model.
- Real editing. OpenAI tested synonym swaps, not editorial rewrites that also restructure sentences.
- Geography beyond the EU. The page says "EU only" and does not mention Switzerland or the EEA.
- Dates. No date is given for the ChatGPT and Codex rollout or for the open-source release of textGrain.
- Error rates outside the approved circle. Nobody without detector access can check how the detector behaves on their own documents.
The opposing case is fair: most misuse is bulk, unedited output, and a mark that catches unedited text catches exactly that. For a regulated document, though, unedited machine text is not the risk case. It is the one case that should never reach a signature.
The better the reviewer, the weaker the mark
The medical writer's job, in every quality system I have worked in, is to take a draft and make it right: change the wording, fix the claim, swap the term for the approved one, cut the sentence that promises too much. That is what human review means. It is also, almost word for word, what removes the mark. The regulation wants machine text to stay recognisable; good practice wants a person to rework it until they can sign it. Do the second properly and you undo the first.
One day a QA reviewer will run a document through a detector her company had to apply for, get a negative result, and learn exactly one thing: that her colleague did a good job. If your quality system is about to meet its first watermarked draft and you want to work out what it should and should not count as evidence, write to me.