OpenAI to watermark text in the EU
On 5 October 2026 OpenAI said how it will meet the EU AI Act's requirement that generated text be machine-readable: API customers worldwide can opt in to text watermarking for select models from that day, invisible watermarks will be added to eligible ChatGPT and Codex text in the EU over the coming weeks, and a detector is open to approved researchers and expert organisations. Its technology, textGrain, adds a statistical signal to word choice.
Why it matters
A company that already marked images and audio now marks text, but only where the law requires it and with an explicit list of what a watermark cannot show. OpenAI's own figures show the limits: shorter or more constrained text is harder to detect, and replacing a quarter of the words lowers detection from about 92% to 17%. The detector stays closed to the public because of the risk of false positives and misses.
What the page says. The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way; OpenAI calls text watermarking and detection "early" technologies with significant limitations. The plan is phased: from 5 October API customers worldwide can opt in to watermarking for select models (off by default); over the coming weeks an invisible watermark will be added to eligible ChatGPT and Codex text for users on all plans in the EU, and only there; the company is also working with cloud partners to make watermarking available in their services; access to the detector is open by application to approved researchers and expert organisations, case by case; the detector will say only whether it detects an OpenAI watermark, without identifying the user or revealing prompts. OpenAI plans to make the technology available as open source. How it works and what was measured. textGrain adds an invisible statistical signal to the model's word choices and the detector looks for it; by OpenAI's own evaluations it matched or exceeded other approaches, SynthID for text included. At a target false positive rate of 1% the detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages (for content such as psychology); for mathematics, where there is less freedom in word choice, detection is substantially lower. Replacing 10% of words with synonyms lowers detection in 400-token passages from about 92% to 66%, replacing 25% to 17%. On the benchmarks OpenAI uses for Astra it sees no meaningful difference with and without the watermark: for example the Artificial Analysis Intelligence Index 49.57 and 49.76 points, GPQA Diamond 94.44% and 93.94%. What a watermark does not tell you. By the page it does not measure human contribution, does not establish ownership or responsibility, does not identify the user, does not verify accuracy, and its absence does not prove human authorship: text may be too short, edited, translated, from an unsupported model or from another company's tools. What the record does not claim. All the figures are OpenAI's own measurements; the technical report on textGrain that the page links to was not read. The page gives no date by which watermarks will begin to appear in ChatGPT and Codex.