OpenAI's EU text watermark cannot tell who did the thinking: schools should redesign assessment

OpenAI announced on 5 October 2026 that it will add an invisible watermark, which it calls textGrain, to text produced by ChatGPT and Codex for users in the European Union. In the company's words, the technology "adds an invisible statistical signal to the model's word choices" rather than inserting hidden characters. It will reach eligible ChatGPT and Codex users on all plans in the EU over the coming weeks; API customers anywhere in the world can opt in for selected models, with watermarking off by default; and access to the detector is initially limited to approved researchers and expert organisations, granted case by case. OpenAI says the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way.

The question teachers, lecturers and managers have asked since ChatGPT arrived is whether a piece of writing came from a person or a model. OpenAI's own figures show why a watermark does not answer it. At a 1% false positive rate its detector identified about 80% of 200-token passages and about 95% of 400-token passages, with lower results for mathematics. Replacing 10% of the words cut detection from 92% to 66%, and replacing 25% cut it to 17%. The announcement also states that a watermark "does not measure human contribution", "does not establish ownership or responsibility", "does not identify the user", and that "the absence of a detected watermark does not prove human authorship".

Dan Fitzpatrick, The AI Educator, says the announcement matters for what it concedes. A watermark detects the tool, not the thinking. A student who drafts with ChatGPT and then rewrites a quarter of it is close to invisible to the detector, while a student who uses it well, to question and improve their own argument, may carry the mark. The thing a teacher needs to know, whether the learner did the thinking, is the one thing OpenAI says the signal cannot measure. Schools, colleges and universities that built their AI policy on detection have now been told by the detector's maker that detection is not the answer. Two further details limit it in practice: the rollout described is for the EU only, and the detector is not being offered to schools or employers.

The practical response is to redesign assessment, not to wait for a better detector. Dan Fitzpatrick suggests leaders start this term: audit every policy and assessment brief for the word "detect" and ask what would happen to it if detection were unavailable, because for most UK institutions it still is. Move the weight of assessment onto the process that can be seen: drafts and version history, short oral explanations of a submitted piece, in-class writing, and tasks that require students to show their reasoning and their sources. Make AI use something students declare rather than hide, and teach them the line that governs it: outsource the doing, not the thinking. Employers should do the same, by judging work on whether the person can explain and defend it, and by setting a clear expectation that AI assistance is declared.

Dan Fitzpatrick works with schools, universities, governments and employers on how AI changes learning and work. He trains teachers and leaders, writes on AI in education, and has spent the period since ChatGPT's launch arguing that the response to AI-written text is better assessment and AI literacy, not an arms race with detection software.

Dan Fitzpatrick is available for interview and comment on this story.