OpenAI adds text watermarking to ChatGPT, Codex in EU first

The Verge reports that OpenAI is introducing an invisible, machine-readable watermark to text outputs from ChatGPT and Codex, initially limited to users in the European Union.

OpenAI calls its system textGrain and says it matched or surpassed rival approaches like Google DeepMind's SynthID for text, which Anthropic cited when it added text watermarking in August. Both companies made the moves to comply with the European Union's AI Act.

OpenAI published benchmark scores showing similar performance between watermarked and unwatermarked text. But the company cautioned that textGrain "does not guarantee reliable detection" and that text watermarks do not verify accuracy, establish ownership, measure human contribution, or prove human authorship.

The rollout to eligible EU users across all plans (free and paid) will happen over the coming weeks. The company is not making it a global default at this time. Starting today, API customers worldwide can opt in to watermarked text outputs for select models. OpenAI is also working with cloud partners to make watermarking available through their services.

Researchers and approved expert organizations can apply for detector access starting today. The detection tool will report whether it finds an OpenAI watermark without revealing the user's identity or their prompts and conversations. OpenAI is not making the detector publicly available because of the risk of missed detections and false positives.

• AI Act
• ChatGPT
• Codex
• EU
• OpenAI
• watermarking