OpenAI details its approach to EU-mandated text watermarking

OpenAI published its approach to complying with the EU AI Act’s requirement that generative AI providers make generated text identifiable in a machine-readable way. The company said its watermarking technology, called textGrain, adds an invisible statistical signal to a model’s word choices that a matching detector can look for; it is rolling the tool out in phases, starting with an API opt-in for select models, followed over coming weeks by watermarking turned on by default for ChatGPT and Codex text output inside the EU, with detector access initially restricted to approved researchers and expert organisations.

OpenAI published its own evaluation of the tool’s limits alongside the rollout. At a 1% false-positive rate, the detector caught watermarks in about 80% of 200-token passages and about 95% of 400-token passages for topics such as psychology, but performed markedly worse on content like mathematics, where word choice offers less flexibility to encode a signal. Editing weakened the watermark sharply: replacing 10% of words in a 400-token passage with synonyms cut detection from roughly 92% to 66%, and replacing 25% cut it to 17%. OpenAI said textGrain matched or exceeded other approaches it tested, including Google’s SynthID for text, but that strong performance under controlled conditions did not guarantee reliable results in everyday use — a caveat it gave as the reason for keeping detector access gated to researchers rather than opening it generally.

The disclosure is notable less for the rollout itself than for how candidly OpenAI quantified a technology it was legally required to ship: the same report that announces the tool also documents, with specific failure rates, how easily a short edit or a technical subject defeats it. OpenAI said it plans to open-source the underlying technique and that the restriction applies only to text; its existing tools for verifying AI-generated images and audio remain public.