Artificial Intelligence
OpenAI plans invisible watermarks for EU ChatGPT and Codex text
OpenAI said it would add invisible watermarks to eligible ChatGPT and Codex output in the European Union, while also offering opt-in watermarking and a detector that could be weakened by editing.
- OpenAI said it would add invisible watermarks to eligible ChatGPT and Codex text output in the European Union over the coming weeks.
- The company said its textGrain method changed word choice enough to create a statistical pattern that could later be detected.
- API developers worldwide could opt in to watermarking for supported models from that day, but the default remained off.
- OpenAI opened applications for a watermark detector and said access would start with approved researchers and expert organisations.
- In OpenAI's tests, replacing 10% of words with synonyms reduced detection on 400-token passages from about 92% to 66%, and replacing 25% reduced it to 17%.
OpenAI said it would add invisible watermarks to eligible ChatGPT and Codex output in the European Union over the coming weeks. The company said the mark would not be visible when text was read or copied.
OpenAI said the system, called textGrain, worked by shifting the model's word choices so that a detectable statistical pattern was left behind. It also said API developers worldwide could opt in to watermarking for supported models from that day, although the default stayed off.
The company also opened applications for a watermark detector. Access would begin with approved researchers and expert organisations rather than the public.
OpenAI said a detected watermark would not identify who generated the text, or reveal the account, prompt or conversation behind it. It also said its tests showed that editing could reduce detection, including synonym replacement in 400-token passages from about 92% to 66% when 10% of words were changed, and to 17% when 25% were changed.
Named in this story
Companies
- OpenAI
- said it would add watermarking and open detector applications
Organisations
- European Union
- the jurisdiction where the watermark would first apply
Products and systems
- ChatGPT
- its text output would receive the watermark in the EU
- Codex
- its text output would receive the watermark in the EU
- textGrain
- the watermarking method OpenAI said it was using
- GPT-6 Astra
- the model whose benchmark results OpenAI cited
How the source tells it
The source read like a technical product update, with caveats and measurements rather than strong emotive framing.
No emotive framing was found in the original.