OpenAI Launches Opt-In Text Watermarking for Select API Models, Plans EU ChatGPT Rollout
Summary
OpenAI launches opt-in text watermarking for select API models worldwide and says it will add invisible watermarks to eligible ChatGPT and Codex outputs in the EU within weeks, while restricting its detector to approved researchers because synonym edits can sharply reduce detection accuracy.
Key Points
- OpenAI launches opt-in text watermarking for select API models worldwide and plans to add invisible watermarks to eligible ChatGPT and Codex outputs in the EU in coming weeks.
- OpenAI limits its text-watermark detector at launch to approved researchers and expert organizations, citing risks of false positives and missed watermarks.
- In tests, replacing 10% of words in 400-token watermarked passages with synonyms cuts detection from about 92% to 66%, while replacing 25% cuts it to 17%.