OpenAI to watermark EU ChatGPT and Codex text with textGrain

OpenAI will invisibly watermark eligible EU ChatGPT and Codex text with textGrain for AI Act rules, while API watermarking stays opt-in worldwide.

OpenAI said it will add textGrain, an invisible statistical watermark, to eligible ChatGPT and Codex text in the EU so generated text is machine-readable under EU AI Act rules; TechCrunch says those transparency rules took effect August 2, and The Verge also cites the Code of Practice. The Decoder and The Verge say the EU change starts within weeks, TechCrunch says it covers eligible users on all plans, and API watermarking on select models stays opt-in, off by default, and available worldwide, including via Microsoft Azure, with TechCrunch saying API customers can enable it today. Ars Technica later says ChatGPT watermarking will be on by default only in the EU and available but off by default elsewhere, a broader availability claim than the other reports' API-focused opt-in. textGrain subtly biases word or token choices so a keyed detector can recognize generated text after copy-paste without identifying the user. At a 1% false-positive target, detection was about 95% for 400-token psychology text and about 80% for 200-token passages, substantially lower for math, and TechCrunch says short passages and translations are harder; replacing 10% of words with synonyms cut 400-token detection from about 92% to 66%, and replacing 25% dropped it to about 17%. OpenAI said quality was essentially unchanged and that textGrain matched or beat SynthID, with The Verge adding Anthropic's watermark, while sources generally cite Astra scores on eight benchmarks including GPQA Diamond and only BleepingComputer says GPT-6 Astra; Ars instead notes that SynthID and C2PA remain relatively easy to circumvent. Detector access starts limited—approved researchers and experts, specialist organizations, or a limited group with an approval process, depending on the source—and OpenAI warns detection is imperfect and a missing watermark does not prove human authorship.

• Eligible ChatGPT and Codex text in the EU will carry invisible textGrain watermarks; The Decoder and The Verge say the rollout is within weeks, and TechCrunch says it covers all plans.
• Worldwide API watermarking on select models is opt-in and off by default, including via Microsoft Azure; TechCrunch says API customers can enable it today, and reports agree it is not a global default at launch.
• Ars Technica says ChatGPT watermarking is on by default only in the EU and the same feature is available but off by default elsewhere.
• TechCrunch says the EU AI Act transparency rules took effect August 2; The Verge also cites the Code of Practice.
• At a 1% false-positive target, detection was about 95% for 400-token psychology text and about 80% for 200-token passages, and substantially lower for math; TechCrunch adds that short passages and translations are harder.
• Replacing 10% of words with synonyms cut 400-token detection from about 92% to 66%, and replacing 25% cut it to about 17%.
• OpenAI reported no meaningful Astra quality drop, including eight benchmarks such as GPQA Diamond; only BleepingComputer calls the results GPT-6 Astra. The Verge says textGrain matched or exceeded SynthID and Anthropic's watermark, while…
• Detector access is initially limited rather than public; sources differ on whether that means approved researchers and experts, specialist organizations, or a limited group plus an approval process. A missing watermark does not prove human…