ChatGPT will watermark text in the EU only. The fine is sized on all of OpenAI's revenue; the mark reaches about 13% of users. Here's the business logic.

OpenAI drew its new text watermark exactly where the law stops, and the map it chose says more about the business than the technology does.

OpenAI will add an invisible statistical watermark, called textGrain, to text from ChatGPT and Codex in the European Union over the coming weeks to comply with the EU AI Act. Everywhere else, consumer ChatGPT text stays unmarked. API customers anywhere can switch the watermark on for select models, but it ships off by default, and the detector that reads the mark is open only to approved researchers for now.

In April 2023, OpenAI asked ChatGPT users a blunt question. Nearly 30% said they would use the product less if it watermarked its text and a rival didn’t. That survey, reported by The Wall Street Journal in 2024, is a big reason a watermark OpenAI already had working sat on a shelf for years.

Three years on, the condition in that question has flipped. Google has watermarked Gemini’s text since 2024. Anthropic started marking Claude’s output worldwide in August. And on Monday, OpenAI switched its own mark on in exactly one region.

Run the arithmetic on that region and a strange shape appears. The fine OpenAI is steering clear of is sized on 100% of its worldwide revenue. The watermark it built to do that reaches roughly 13% of ChatGPT’s weekly users. Outside Europe, OpenAI has become the rival from its own survey.

OpenAI laid out the plan on October 5 in a post titled Our approach to EU text provenance rules, and TechCrunch reported it the same day. The rollout has three parts, and each one ships with a different default:

The method, textGrain, nudges the model’s word choices using a secret key. The result is a statistical pattern a reader can’t see and a detector holding the key can. Because the signal lives in the words themselves, it survives copy and paste. OpenAI published a technical report co-written with researchers from the University of Pennsylvania and Yale, says it plans to open-source the technology, and claims textGrain matched or beat every alternative it tested, Google’s SynthID included.

The company was unusually frank about the limits. At a 1% false positive target, its detector caught about 80% of 200-token passages and about 95% of 400-token passages on psychology questions, with math answers scoring far lower. Swap 10% of the words in a 400-token passage for synonyms and detection falls from about 92% to 66%. Swap 25% and it lands at 17%.

Quality held up. On OpenAI’s benchmark table for Astra, its newest frontier model, watermarked and unwatermarked scores sit within about a point of each other on most tests: 94.44% versus 93.94% on GPQA Diamond, 87.92% versus 87.35% on BrowseComp. On a couple of tests the watermarked version actually scored higher.

Two more lines from the post matter. The mark does not identify the user, and a missing mark does not prove a human wrote the text.

Article 50 of the EU AI Act took effect on August 2, 2026. It requires providers of generative AI systems to mark synthetic output in a machine-readable, detectable way. Systems already on the market before that date have until December 2 to comply, and providers need a detection interoperability solution by February 2, 2027. Under Article 99(4), breaking Article 50 can cost up to €15 million or 3% of total worldwide annual turnover, whichever is higher.

OpenAI signed the EU’s Code of Practice on AI-generated content alongside Anthropic, Google, Meta and Microsoft. Its text watermark, though, is older than the rule. According to the Journal’s 2024 reporting, the system had been debated internally for about two years and was 99.9% effective when enough new text was generated. Two surveys from April 2023 pulled in opposite directions. People worldwide backed an AI detection tool four to one. Among ChatGPT users, 69% believed cheating detection would produce false accusations, and nearly 30% said they would use ChatGPT less if it watermarked and a rival didn’t.

Then the rivals moved. Google DeepMind published SynthID-Text in Nature in October 2024 and reported it was live across Gemini, after a test on nearly 20 million responses found no meaningful difference in user ratings. Anthropic said on August 11 that Claude would mark text worldwide, across its API, its apps and the big cloud marketplaces, which we covered when it landed. That choice cost something. Within a day, TechCrunch was reporting on Claude users angry that work they had directed would carry a machine’s signature.

One correction to our own reading. In August we argued Anthropic went global because one inference path costs less than two. Anthropic later said the worldwide rollout reflected the lack of a durable way to scope marking by region. OpenAI has now run the other experiment and concluded that two paths are worth paying for.

Stack the three switches on top of each other and the design principle is clear. OpenAI marks what OpenAI ships, in the one jurisdiction that requires it, and offers a switch for everything else.

That tracks the law closely. Article 50(2) puts the marking duty on the provider of the AI system. ChatGPT is OpenAI’s system, so ChatGPT gets the mark in the EU. A company that builds its own app on the API is, on a plain reading, the provider of that app. OpenAI’s post says the opt-in lets customers decide how watermarking fits their transparency obligations. Translation: the decision and the duty move downstream together.

The detector gate finishes the design. OpenAI cites the Code of Practice for its case-by-case access and the risk of misses and false positives for keeping the tool closed. Both reasons hold up. The commercial side effect is real too. A watermark only changes behavior when someone other than the issuer can read it, which is the whole reason AI detection became a business in the first place. For text, the only readers right now are OpenAI and the researchers it approves.

Compare that with images and audio. OpenAI already runs a public verification tool at openai.com/verify for both. Where the signal is robust, the company lets anyone check. Where it is fragile, it holds the key.

Start with coverage. In its Digital Services Act disclosure, OpenAI reported about 159.1 million average monthly active recipients of ChatGPT search in the EU for the six months to March 2026, the figure that got ChatGPT designated a Very Large Online Search Engine. At DevDay on September 29, OpenAI cited 1.2 billion weekly users. Divide one by the other and you get 13.3%.

Treat that as an estimate. Monthly counts run above weekly ones, which pushes the true weekly EU share down, while six months of growth since March pushes it up. Call it roughly one ChatGPT user in eight.