OpenAI introduces invisible statistical watermarking for ChatGPT texts in the EU

Summary

OpenAI is rolling out 'textGrain,' a statistical word-selection-based watermark for ChatGPT and Codex outputs in the EU. This move aims to comply with European AI transparency regulations. The technique creates a detectable pattern without adding visible or hidden characters. However, OpenAI acknowledges that the watermark can be weakened by minor edits and is unreliable for short texts.

Why it matters

OpenAI has become the first major US lab to adopt a text-specific provenance signal in the EU, following a similar step by Anthropic in August. This action provides a practical template for how AI-generated text can be marked under the EU's transparency obligations. It also highlights a structural gap: while C2PA/SynthID are already functional for images and audio, text watermarking remains probabilistic and easily disrupted. This will shape how regulators and platforms handle 'AI content' labels in the coming years.

Highlights

• textGrain works by statistically guiding word choice rather than adding hidden characters or visible markers.
• In a 400-token text, replacing just 10% of words with synonyms drops detection accuracy from 92% to 66%.
• The rollout begins in the EU as a real-world pilot. API users can enable this feature globally on selected models.

Critical take

OpenAI's own data undermines the core assumption of this approach. A 25% rate of synonym substitution reduces detection rates to 17%. In short texts, the watermark remains close to noise levels. This suggests the watermark functions more as a regulatory checkbox than a robust provenance tool. The system cannot distinguish between AI-assisted text and fully AI-generated text, nor can it identify the user or prompt. Consequently, the fundamental transparency question targeted by EU regulations remains largely unanswered.