signalAI热榜2026-10-06
OpenAI Unveils Text Provenance Solution Under EU AI Act, Launches textGrain Text Watermark
OpenAI has unveiled textGrain, a text watermarking technique that embeds an invisible statistical signal into word choices, to comply with EU AI Act provenance rules. API customers can now opt in for select models, and invisible watermarks will be added to ChatGPT and Codex outputs in the EU within weeks, while detector access is initially limited to approved researchers. Evaluations show detection rates of about 80% for 200-token passages and 95% for 400-token passages, but they drop sharply for low-flexibility content like math and after heavy editing, with 10% synonym replacement reducing detection from about 92% to 66% and 25% to 17%.
- for who
- OpenAI API customers, EU ChatGPT and Codex users, and researchers working on text provenance
- why now
- EU text watermarking launches now for APIs, EU ChatGPT and Codex within weeks.
- what changes
- API customers gain an optional watermarking feature, EU ChatGPT and Codex outputs will carry invisible watermarks, and qualified researchers can apply for detector access
- to do
- API customers can enable text watermarking for select models today, and researchers can apply for detector access
key points
- textGrain embeds an invisible statistical signal into model word choices for provenance
- Detection: about 80% for 200 tokens, 95% for 400 tokens, lower for math and edited text
- API opt-in available now, EU watermarks in weeks, detector restricted to approved researchers
#text watermark#eu ai act#openai#textgrain3 sources · confidence medium
score
score 8 out of 10. 0-10: how dense the facts are, multiplied by how much you can do with them after reading. 8+ means the topic's evidence bar is met: benchmarks and availability for a new model, amount and investors for a funding round, revenue figures for a solo-money story. Below 5 an item does not enter the digest. A press release scores 3 or less, a reprint loses 2, anything older than 14 days loses 1, a headline that misleads loses 3.
read the source