OpenAI explains its approach to EU text provenance rules

OpenAI said it is introducing text watermarking and related provenance tools to help identify AIgenerated text in line with the EU AI Act. The company said the rollout will start with optional watermarking for some API models globally, followed by invisible watermarks for eligible ChatGPT and Codex text output in the European Union over the coming weeks. The company described its text watermarking system, called textGrain, as an invisible statistical signal added through word choices. OpenAI said its detector can find the signal in some cases, but performance varies by text length, subject matter, and editing, and false positives and false negatives remain possible. OpenAI said detector access will initially be limited to approved researchers and expert organizations, while image and audio provenance tools will remain publicly available. The company also said it does not see meaningful quality changes from watermarking in its benchmark testing and plans to continue improving the technology as standards and regulations evolve.