OpenAI is adding text watermarking in ChatGPT and Codex

OpenAI has introduced a new text watermarking system called textGrain for its API, allowing developers to opt-in to this feature. The system subtly influences the words a language model chooses, creating a pattern detectable by OpenAI's detector over long enough passages. API customers worldwide can enable text watermarking on supported models starting today, with automatic watermarking for ChatGPT and Codex text in the EU beginning in the coming weeks.

This response follows EU AI Act transparency requirements. Unlike Anthropic, OpenAI's watermarking is off by default in the API, giving customers control over its implementation. Text watermarking remains off by default in the API to let customers decide how it fits their transparency obligations and user experiences. OpenAI's detection performance for textGrain matches or exceeds SynthID, with plans to open-source the technology.

However, the technique has limitations, such as reduced detection rates for shorter passages and edited content, and difficulty watermarking code due to limited plausible choices for subsequent words. OpenAI plans to automatically watermark eligible Codex text output in the EU, acknowledging the challenges in watermarking source code.