OpenAI starts text watermarking for the EU AI Act

OpenAI published its approach to text watermarking, framed as a response to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine readable way. The rollout is deliberately phased in three parts. Starting today, API customers globally can opt in to text watermarking for select models, and it stays off by default in the API. Over the coming weeks an invisible watermark will be added to eligible ChatGPT and Codex text output in the European Union only. And applications open for access to the text watermark detector, initially limited to approved researchers and expert organizations.

The technology is called textGrain, and it adds an invisible statistical signal to the model’s word choices. OpenAI says it matched or exceeded the other approaches it tested, including SynthID for text, plans to publish a technical report, and intends to make the technology open source.

OpenAI spends more of the post on limits than on capabilities, and the numbers are specific. At a target false positive rate of 1%, the detector found watermarks in about 80% of 200 token passages against about 95% of 400 token passages for content like psychology, and detection was substantially lower for mathematics where there is less flexibility in word choice. Editing degrades it further: replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% dropped it to 17%. On output quality the company reports no meaningful performance differences on Astra with and without watermarking, listing for example Artificial Analysis Intelligence Index at 49.57 unwatermarked against 49.76 watermarked, and DeepSWE v1.1 at 72.80% against 71.68%.

The post is explicit about what a detection result does not establish: it does not measure human contribution, does not establish ownership or responsibility, does not identify the user, does not verify accuracy, and the absence of a watermark does not prove human authorship, since text may be too short, edited, translated, generated before watermarking, or produced by another company’s tools. Given the risk of both missed watermarks and false positives, the detector is not public at launch, and the EU rollout is not a global default.

(Original source: OpenAI)