OpenAI unveils text watermarking for EU AI Act

Summary

OpenAI shared its approach to EU AI Act text provenance rules with a statistical watermarking method called textGrain. API customers worldwide can opt in, and an invisible watermark will be added to eligible ChatGPT and Codex text output in the EU in the coming weeks. In the company's own evaluations at a 1% false positive rate, the detector caught about 80% of 200-token passages and 95% of 400-token passages, but the signal weakens under edits such as synonym replacement. Detector access is initially limited to approved researchers and organizations, and an open-source release is planned.

Why it matters for our work

As regulators push machine-readable identification of AI-generated text, provenance marking may become standard practice in research and document work. The reported detection limits, however, should be factored into any policy or workflow decision.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.