OpenAI details invisible text watermarking for ChatGPT, Codex and API

OpenAI announced textGrain, an invisible text-watermarking system intended to help identify material generated by its models. The company presented the system in the context of the EU AI Act’s requirements for identifying AI-generated content.

API customers globally can opt in to watermarking for selected models starting today. OpenAI plans to add it to eligible ChatGPT and Codex text outputs in the European Union in the coming weeks.

How textGrain works

textGrain adds an invisible statistical signal to a model’s word choices. A detector searches for that signal and assesses whether a passage contains an OpenAI watermark.

OpenAI has opened applications for detector access. Initially, only approved researchers and expert organizations will be able to use it while the company evaluates and improves the technology. OpenAI also plans to release the system as open source.

Detection limits

OpenAI warns that the detector can produce false positives and false negatives, particularly for short passages and domains with less flexibility in word choice, such as mathematics.

In a test of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. Editing and translation can also weaken the watermark.

The watermark does not measure human contribution, establish ownership or responsibility, identify the user, or verify accuracy. OpenAI also says that an undetected watermark does not prove human authorship.

The company said it will revisit the approach as the technology, standards, and evidence develop, including research into whether watermarking can better distinguish AI assistance from AI authorship.