OpenAI Will Watermark ChatGPT’s Words in Europe — But Not Anywhere Else

BRUSSELS — OpenAI is finally putting its invisible ink to use. The company announced Monday that it will start embedding a statistical watermark into text generated by ChatGPT and Codex for users in the European Union — a quiet, EU-only move designed to satisfy the EU AI Act’s transparency requirements, and one that leaves everyone else in the world exactly where they were.

What textGrain actually does

The watermark, which OpenAI calls textGrain, is not a visible symbol or hidden metadata. It works by subtly shaping the model’s word choices as it generates text: using a secret key, the system sorts the model’s next-word predictions so that hundreds of tiny nudges across a passage add up to a pattern a detector can recognize — using only the text and the key. A reader sees nothing unusual. Because the pattern lives in the words themselves, it survives copy and paste, which is exactly where metadata watermarks die.

OpenAI published a technical report on the method alongside the announcement, co-written with researchers from the University of Pennsylvania and Yale, walking through the key-and-detection math in detail. Two properties the company emphasizes: the watermark does not identify the user who generated the text — it marks the text as machine-shaped, not the person behind it — and OpenAI says it observed no meaningful change in model performance with the watermark switched on.

Who gets it, and who doesn’t

The rollout covers eligible ChatGPT and Codex users on all plans — but only inside the EU — arriving over the coming weeks. Separately, developers using OpenAI’s API anywhere in the world can switch the watermark on for select models starting now; it is off by default. OpenAI is explicit that text watermarking will not be a global default at launch.

That asymmetry is the story. OpenAI is not doing this because it fell in love with watermarking; it is doing it because the EU AI Act’s transparency rules — Article 50 — require AI companies to mark AI-generated content in a way other systems can identify. The rules took effect on August 2, 2026, and systems that already existed before then must comply by December 2, 2026. Non-compliance carries fines of up to €15 million or 3% of global turnover — the kind of number that gets a general counsel’s attention even at OpenAI’s scale. Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to following the EU’s code of practice on AI-generated content.

The contrast with Anthropic is deliberate-looking. Two months ago Anthropic said it would watermark text generated by Claude — and applied the decision worldwide. That move drew backlash from some Claude users, who argued they had supplied “the instructions, context, decisions” while Claude was just “the tool.” OpenAI’s narrower approach neatly avoids that fight at home while ticking the Brussels compliance box. For more on how the two labs are diverging, see Anthropic’s Claude Sonnet 5.5 launch and OpenAI’s shelved GPT-6.1 Astra over safety concerns.

What it can’t do: the cheating question

Here is the part OpenAI does not want misunderstood. The company’s own tests show the watermark can be weakened by ordinary editing. In one test, replacing just 10% of a passage’s words with synonyms dropped detection from about 92% to 66%. Short passages, math answers, and translated text are also harder to detect. And a missing watermark, OpenAI cautions, “does not prove human authorship” — the text could be too short, too heavily edited, or generated by another company’s AI altogether.

“Watermarks can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” the company said. That caveat matters for the most-requested use case — catching students and cheaters — because a 66%-detection system with false-positive risk is a disciplinary nightmare, not a solution. OpenAI’s own history backs the caution: the company built a text watermark before and held off on releasing it, partly over concerns that users would defect to rivals that didn’t watermark — a fear The Wall Street Journal reported on in 2024.

That history also explains the detector strategy. “These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” OpenAI said. The detector is not launching as a public cheating-detection tool; it is starting as a research instrument, with reliability still an open question.

What to watch

The immediate question is how smoothly the EU rollout goes — whether the watermark measurably degrades anything in real use (OpenAI says no), and whether European users even notice it. The bigger question is whether “EU-only” holds. OpenAI left the door open with the API opt-in; if the research community validates the detector and the user backlash Anthropic absorbed doesn’t materialize for a quieter default, the company that once shelved watermarking for fear of losing users could end up making it standard. For now, though, the message is calibrated: comply in Brussels, reassure everywhere else, and don’t promise anyone a cheater-detector.