Technology
OpenAI plans invisible watermarks on European Union text to satisfy regulations
The company will apply its textGrain system to European ChatGPT and Codex outputs while offering global API opt-ins.
The short version
- OpenAI will begin embedding invisible watermarks in eligible ChatGPT and Codex text in the European Union to follow EU AI Act transparency rules.[Hacker News]
- The textGrain technology embeds statistical signals into word selection, and global API users can voluntarily opt in on select models.[Engadget · Hacker News]
- Detection tool access is initially limited to verified researchers and specialist groups amid known accuracy drops from short length or text editing.[Engadget · Hacker News]
Key facts
- The EU AI Act mandates that generative AI providers ensure generated text can be identified in a machine-readable format.[Hacker News]
- OpenAI uses a proprietary system named textGrain to add an invisible statistical signal to model word choices.[Engadget · Hacker News]
- OpenAI will add invisible watermarking to eligible European Union ChatGPT and Codex outputs over the coming weeks.[Hacker News]
- Worldwide API customers can opt in to text watermarking on select models immediately.[Hacker News]
- Access to OpenAI's watermark detector is initially restricted to approved researchers and expert organizations.[Engadget · Hacker News]
What remains uncertain
- Detector reliability varies by passage length, identifying about 80% of 200-token passages versus roughly 95% of 400-token samples at a 1% false positive rate.[Hacker News]
- Watermark detection drops significantly under text modification, falling from about 92% to 66% in 400-token texts when 10% of words are replaced with synonyms.[Hacker News]
Sources
Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.