OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex for users in the European Union, the company said on Monday (5 October, European time) in a blog post on its EU text provenance approach. According to TechCrunch, the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans — but only in the EU, where the AI Act’s transparency obligations took effect on 2 August and require AI-generated content to be marked in a way other systems can identify. Developers using the API anywhere in the world can already switch watermarking on for select models; it is off by default, and OpenAI is not making it a global default at launch.
The system is called textGrain. Rather than stamping a symbol into the text, it subtly shapes which words the model picks — nudging low-stakes choices between near-synonyms in a pattern readers cannot see but a detector holding the secret key can pick up, even after the text is copied and pasted. OpenAI co-wrote a technical report on the method with researchers from the University of Pennsylvania and Yale, and The Verge’s write-up notes the company found no meaningful performance change with the watermark switched on, and says it does not identify the user.
The caveats are doing a lot of work
Three limitations ship alongside the watermark, and each blunts its practical power. First, editing defeats it: in OpenAI’s own test, replacing just 10% of words with synonyms dropped detection from about 92% to 66%. Short passages, math answers and translated text are harder to detect for the same structural reasons. Second, detector access is deliberately narrow — restricted to approved researchers and expert organisations, who can help evaluate reliability and responsible use, rather than a public tool. Third, OpenAI cautions the absence of a watermark does not prove human authorship, since text could be too short, too heavily edited, or generated by a different company’s model. As the company put it, a watermark “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.”
Unite.AI’s phased-rollout read, republished by TechAIwire, frames the launch in three rows: API customers worldwide can opt in from 5 October; EU ChatGPT and Codex users get it automatically on every plan in the coming weeks; and researcher detector applications opened the same day. Users outside the EU get nothing under this plan — a deliberate geography, not a technical limit.
We have seen this playbook before — twice
The launch confirms a pattern this site has tracked since August. When Anthropic committed its future Claude models to an invisible SynthID-Text watermark, the mechanism was the same: statistical shaping of word choice, invisible to readers, checkable only with the key — a method Google DeepMind has run on Gemini since 2024. As we wrote at the time, the EU AI Act has now split AI into two camps: cloud providers that confess their provenance, and open models that carry no watermark at all. OpenAI joining means the three most-used proprietary writing assistants on the planet all sign their work in the EU — while anyone running an open model locally leaves no trace by design.
The geography is the more interesting leak. OpenAI could have gone global; it chose jurisdictional minimum compliance instead. That tells you what the watermark is: not a safety feature, a regulatory checkbox. New Zealand sits on the wrong side of that line — ChatGPT text for Kiwi users carries no watermark under this plan, and there is no NZ equivalent of the AI Act’s Article 50 transparency rule in force. For teachers and editors here, the practical effect of the EU’s move is zero today: the detector is restricted, the watermarked text is not flowing to non-EU accounts, and proving a passage is AI-written remains as inconclusive as it was last week. Regulation by market size means EU rules become the world’s rules at the pace of commercial convenience — and for watermarking, “coming weeks” in Brussels did not come with a ship date anywhere else.
Two open questions worth tracking: whether academic publishers — who have driven the AI-detection demand since ChatGPT’s release — get approved detector access at scale, and whether OpenAI’s numbers survive contact with adversarial paraphrase tools, which exist precisely to defeat statistical watermarks.
