OpenAI has announced a phased rollout of text watermarking to meet EU AI Act provenance requirements. Starting today, API customers globally can opt in for select models, while eligible ChatGPT and Codex users in the EU will receive watermarked output in the coming weeks. The feature remains off by default in the API.
The technology, called textGrain, embeds an invisible statistical signal in word choices. OpenAI says it matched or exceeded other approaches like SynthID for text, but detection is less reliable for short or constrained text, and synonym replacement can sharply reduce detection rates. That is why detector access is initially limited to approved researchers and expert organizations.
OpenAI reports no meaningful performance differences on Astra benchmarks with and without watermarking. The company also stresses that a watermark does not measure human contribution, establish ownership, identify users, or verify accuracy, and the absence of a watermark does not prove human authorship.