OpenAI textGrain watermarking for the EU

Ars Technica on October 7 placed textGrain, OpenAI's statistical text watermark for EU ChatGPT and Codex outputs, alongside SynthID and C2PA, characterizing all three as relatively easy to circumvent with basic know-how, and confirmed that OpenAI has a request-for-approval process for researchers and organizations seeking detector access beyond the initial approved set.

The Neuron Daily on October 6 added OpenAI's statement that a watermark cannot show who wrote a passage, how much a human edited it, or whether it is accurate.

OpenAI announced on October 5 that it is rolling out text watermarking for ChatGPT and Codex outputs in the EU to comply with Article 50 of the EU AI Act, which requires machine-readable marks on generative AI outputs.

The watermark, called textGrain, embeds a hidden signal by statistically tilting word choices, extending provenance tools already applied to images and audio. OpenAI's own figures show the detector caught roughly 80% of 200-token passages at a 1% false positive rate, with detection dropping from around 92% to 17% when 25% of words in 400-token passages are replaced with synonyms. Access to the detector is restricted to OpenAI and approved researchers; API customers globally can opt in starting October 5, while EU deployment of ChatGPT and Codex watermarking will proceed over the coming weeks.