On Monday, October 5, 2026, OpenAI announced on its blog that it will add an invisible watermark to text generated by ChatGPT and Codex inside the European Union. The company framed the move as its way of meeting the transparency obligations of the EU Artificial Intelligence Act (AI Act), which requires generative AI providers to mark their output so that other systems can identify it automatically.
According to reporting by TechCrunch and Tech Times, the technology is called textGrain. The rollout will be gradual: over the coming weeks it will reach eligible ChatGPT and Codex users on every subscription plan, but only within the EU. OpenAI made clear that it is not making text watermarking a global default at this initial launch and that it wants real-world experience before considering any wider expansion.
For developers the situation is different. Starting on October 5, any OpenAI API customer, in any country, can turn the watermark on for selected models. The option is off by default, so companies that build products on top of OpenAI's models will have to decide explicitly whether they want their text to carry the signal.
The watermark is not a visible symbol or a piece of metadata appended to the end of the text. It works by subtly shaping the model's word choices, leaving a statistical pattern that readers cannot perceive but that a detector can recognize. Because the signal lives in the words themselves, it travels with the text when it is copied and pasted into another document. OpenAI says the watermark does not identify the user and that it saw no meaningful change in its models' performance with it switched on.
Alongside the announcement, the company published a technical report on textGrain, co-written with researchers from the University of Pennsylvania and Yale University. The paper explains how a secret key is used to sort next-word predictions. Each choice adds a small nudge; when hundreds of those nudges are added together, the detector can recognize AI-generated content using only the text and the key.
OpenAI was candid about the method's limits. In one of its tests, replacing 10% of the words with synonyms reduced the detection rate from about 92% to 66%. The company added that short passages, math answers and translated text are harder to detect. In other words, moderate editing can weaken the signal considerably, and translation can largely erase it.
Because of these limitations, OpenAI decided that initial access to the detector will be restricted to approved researchers and expert organizations, who will help evaluate its reliability and responsible uses. The company also warned that a missing watermark “does not prove human authorship”: the text may be too short, heavily edited or produced by another company's system.
Another important caveat concerns authorship. According to OpenAI, a watermark can indicate that one of its systems generated or processed part of a passage, but not how much human judgment, editing or creativity went into it. That nuance speaks directly to the debate over AI use in schools, journalism and professional work, where a positive detection could be interpreted unfairly.
The announcement comes two months after Anthropic said it would watermark text generated by Claude, in its case worldwide. That decision drew criticism from some of its users, who argued that they supplied “the instructions, context, decisions” while Claude was just “the tool.” OpenAI has chosen a narrower path: mandatory in the EU, optional on the API and, for now, absent for ChatGPT users in the rest of the world.
The history of this technology inside OpenAI is not new. In 2024, The Wall Street Journal reported that the company had already built a text watermark but held off on releasing it, partly over fears that users would move to competitors that did not mark their output. The European legal requirement changes that calculation: if every major provider must comply in the EU, the competitive disadvantage shrinks inside that market.
The regulatory framework also explains the timing. The AI Act's transparency rules took effect on August 2, and Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that committed to follow the EU code of practice on AI-generated content. textGrain is therefore OpenAI's concrete answer to a commitment it had already made publicly.
For European users and organizations, the immediate practical effect will be limited: writing in ChatGPT will not change visibly, and detection will only be in the hands of a small group of evaluators. Over the medium term, however, universities, media outlets, platforms and public administrations could rely on these detectors for transparency policies, which makes it essential to remember their error rates and their fragility under editing.
In conclusion, textGrain is a meaningful step toward traceability for AI-generated text, but it is not definitive proof of authorship. OpenAI complies with Brussels, limits the geographic scope and openly acknowledges the weaknesses of its method. The big question for the coming months is whether regulators, schools and companies will use these signals as useful indicators rather than automatic verdicts.
