OpenAI Is Watermarking ChatGPT Text in the European Union

OpenAI announced on October 5, 2026, that it will add invisible watermarks to text generated by ChatGPT and Codex in the European Union to comply with transparency requirements under the EU AI Act.

The proprietary system, named textGrain, embeds an invisible statistical signal into the model's word choices. Co-developed with researchers from Yale University and the University of Pennsylvania, the technology is designed to preserve text quality. In internal evaluations, OpenAI reported output scores of 49.57 points for watermarked text compared to 49.76 points for unwatermarked text.

How textGrain Works and Its Limitations

Statistical watermarking has clear technical limits when text is altered. In OpenAI's evaluations of 400-token passages, replacing 10 percent of words with synonyms reduced detection accuracy from approximately 92 percent to 66 percent. Replacing 25 percent of words dropped accuracy to 17 percent.

Reliability also depends on subject matter. Tested at a 1 percent target false-positive rate, textGrain reached an 80 percent detection rate on 200-token psychology passages and 94 percent to 95 percent on 400-token psychology passages. By comparison, 400-token mathematics responses achieved roughly 60 percent accuracy, where technical constraints leave fewer alternative word choices.

OpenAI cautioned that the absence of a detected watermark does not prove human authorship. The company also stated that the watermark does not identify the user, account, prompt, or conversation.

Geographic Rollout and API Settings

The rollout follows the EU AI Act's Article 50 transparency requirements, which took effect on August 2, 2026. While Anthropic deployed text watermarking globally across Claude models, OpenAI has confined mandatory watermarking to EU ChatGPT and Codex instances across all subscription tiers over the coming weeks.

For global API developers, text watermarking is available on select models as an opt-in feature that remains off by default. OpenAI plans to expand opt-in API watermarking support to partner cloud platforms, including Microsoft Azure, in the coming weeks.

Restricted Access to Detection Tools

Access to detection remains narrow. OpenAI is not releasing its detector to the general public, instead restricting access to vetted researchers and expert organizations on a case-by-case basis.

OpenAI claimed textGrain matched or exceeded existing watermarking methods it evaluated, including Google DeepMind's SynthID for text. While the company stated that it plans to eventually release textGrain as open source, it has not published an exact release date or confirmed whether public detection tools will follow.