OpenAI to watermark ChatGPT text in EU
OpenAI says it will add invisible watermarks to ChatGPT and Codex text in the EU to comply with the AI Act, while warning that short or edited passages are hard to detect reliably.
OpenAI has published its plan for meeting the EU AI Act requirement that AI-generated text be identifiable by machines. The plan has three parts:
- EU default for ChatGPT and Codex. Over the coming weeks, eligible text output on all plans will carry an invisible watermark, in the EU only. OpenAI said it is not making watermarking a global default at launch.
- Opt-in for the API. Starting today, API customers worldwide can turn on watermarking for select models. It stays off by default. OpenAI is also working with cloud partners to offer it in the coming weeks.
- Limited detector access. Approved researchers and expert organizations can apply to use the detector. It is not public.
The technology, called textGrain, adds a statistical signal to the model's word choices. OpenAI says textGrain matched or exceeded the other methods it tested, including Google's SynthID for text. A technical report is out now, and OpenAI plans to release textGrain as open source.
OpenAI also reported limits. At a 1% false-positive rate, the detector caught about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. Detection rates for math were "substantially lower." Replacing 10% of words with synonyms in 400-token passages cut detection from about 92% to 66%. Replacing 25% cut it to 17%.
OpenAI says watermarking did not meaningfully affect output quality on benchmarks for Astra, its latest frontier model. For example, GPQA Diamond scored 94.44% without the watermark and 93.94% with it.
The company stressed that a detection result does not identify the user, measure human contribution, or confirm accuracy. A missing watermark also does not prove a human wrote the text. OpenAI's existing image and audio verification tools remain publicly available.