OpenAI will watermark ChatGPT and Codex text in the EU: what textGrain means for enterprises

OpenAI published its approach to the EU's text provenance rules on 5 October 2026. Over the coming weeks it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union, for users across all plans, in the EU only. OpenAI says it is not making text watermarking a global default at launch, and that the regional approach gives it room to learn from real use.

The technology is called textGrain. OpenAI describes it as an invisible statistical signal added to the model's word choices, which a detector can then look for. It has published a technical report, says it will update it in the coming weeks and plans to release the technology as open source.

OpenAI ties the change to the EU AI Act, which it says requires generative AI providers to make generated text identifiable in a machine-readable way. The European Commission's guidance on Article 50 says providers must mark AI-generated audio, image, video and text in a machine-readable format so that it can be detected as artificially generated. Article 50 has applied since 2 August 2026. For systems placed on the market before that date, the marking and detection obligation applies from 2 December 2026.

A company with staff on both sides of the EU border will now produce two kinds of ChatGPT text: marked where it was generated in the EU, unmarked elsewhere. Nothing visible changes for the person typing, and OpenAI describes no setting for it in its announcement. The practical effect is that documents, emails and code comments written with ChatGPT in an EU office can later be tested for an OpenAI watermark by whoever holds detector access.

OpenAI is clear about what the mark does not carry. It says a watermark does not identify the user, and does not associate a person, organisation, account, prompt or conversation with the text. On OpenAI's account, that limits the privacy concern, and it also means the mark cannot be used to trace a leak to an employee.

On quality, OpenAI reports no meaningful performance difference with and without watermarking across the benchmarks it uses for its Astra model, and publishes the paired scores. Those are OpenAI's own evaluations. A team with a sensitive workload, such as structured output or code, can run its own comparison before assuming the same.

From 5 October 2026, API customers around the world can opt in to watermarked text outputs for select models. Text watermarking remains off by default in the API. OpenAI says this lets customers decide how watermarking fits their own transparency obligations and the experiences they provide to users.

OpenAI also says it is working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.

This is the part that lands on businesses. A company that builds an assistant, a drafting tool or a support bot on the API and offers it under its own name is the one that decides whether its text is marked. The Commission's guidance defines a provider as a body that develops an AI system, or has one developed, and places it on the EU market or puts it into service under its own name or trademark, wherever that body is established. A UK or US company whose product's output is used in the EU can fall within that.

The same guidance sets out exceptions and a narrow exemption: for example where the system only assists with standard editing, and for outputs used in business-to-business or industrial contexts where the conditions in the Commission's guidelines are met. Whether a given product is in scope is a legal question. OpenAI's switch is one way to meet a marking obligation where there is one, and the dates above set the timetable. By our count, 2 December 2026 is under eight weeks from the date of this post.

The Commission also describes a voluntary Code of Practice on Transparency of AI-generated Content, which it and the AI Board have assessed as adequate for showing compliance, and says about 190 companies and organisations had signed it by the end of July 2026. Companies that comply by other means have to show those means are adequate. The guidance says fines can reach 15 million euros or 3% of total worldwide turnover.

None of this is legal advice. It is a prompt to put the question to counsel now, with the product list in hand.

OpenAI has opened applications for access to its text watermark detector. Access is at first limited to approved researchers and expert organisations, granted case by case. The tool reports whether it detects an OpenAI watermark, without identifying the user or revealing prompts or conversations. OpenAI says that, given the risk of missed watermarks and false positives, it is not making the detector public at launch.

It publishes the reasons. At a target false positive rate of 1%, its detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, with substantially lower rates for content such as mathematics, where there is less freedom in word choice. In a test on 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

OpenAI lists what a detection result cannot show. A watermark does not measure how much a person contributed, establish ownership or responsibility, identify the user or verify accuracy. And a missing watermark does not prove a person wrote the text: it may be too short, edited or translated, come from an unsupported model, predate watermarking or come from another company's tools.

For an employer, a university or a publisher, the conclusion follows from OpenAI's own numbers. A result that falls from about 92% to 17% after a quarter of the words are swapped is unlikely to carry a disciplinary case, an academic misconduct finding or a contract dispute. OpenAI's decision to withhold the detector from the public is consistent with that.

It also means a business cannot check its own suppliers' or candidates' text for an OpenAI watermark today. Organisations that have bought third-party AI text detectors should note that the announcement does not give those tools access to the detector.

What the watermark does do is let a provider show a regulator that its output is marked. That is a compliance function. It is a different thing from a tool for deciding who wrote a document, and policies are safer when they keep the two apart.