OpenAI introduces textGrain watermarks for ChatGPT and Codex in the EU, optional for API globally

On October 5, OpenAI published its approach to EU text provenance rules under the EU AI Act, formally launching textGrain watermarking technology. textGrain is an invisible statistical signal added to model word choices, allowing detectors to identify whether text carries the OpenAI watermark. OpenAI also committed to open-sourcing textGrain for other developers.

On the rollout, OpenAI announced three steps: first, eligible ChatGPT and Codex text outputs in the EU will receive invisible watermarks by default in the coming weeks; second, API customers worldwide can opt in for select models starting today; third, detector access applications are open to approved researchers and expert organizations. Watermark capability will also be progressively extended to OpenAI model outputs served through cloud partners.

At a 1% false-positive target, textGrain detects around 80% of 200-token passages and around 95% of 400-token passages; after 25% synonym replacement the rate drops to 17%, and after 10% replacement from 92% to 66%. Mathematical content sees significantly lower detection rates because of reduced vocabulary flexibility. OpenAI benchmarked watermarked vs unwatermarked versions of Astra and similar models on Artificial Analysis Intelligence Index, AutomationBench, DeepSWE v1.1 and Terminal-Bench 4.0, with differences within one percentage point and no significant impact on output quality.

OpenAI is explicit about what watermarks cannot do: measure human contribution, establish ownership or accountability, identify specific users and accounts, or verify factual accuracy. The company stresses that the absence of a watermark does not mean human authorship — short text, editing, translation, or use of another company's tool can also produce unmarked text. The approach also adopts Content Credentials content provenance, embedding SynthID watermarks in supported image and audio outputs.