OpenAI Details EU Text Provenance Approach and textGrain Watermarking

OpenAI has shared its approach to text watermarking in response to the EU AI Act, introducing a phased rollout of its textGrain technology. Starting October 5, 2026, global API customers can opt in to text watermarking for select models, though the feature remains off by default. Over the coming weeks, the company will add an invisible watermark to eligible ChatGPT and Codex text output specifically within the European Union.

The technology, called textGrain, embeds an invisible statistical signal into a model's word choices. OpenAI is also opening applications for approved researchers and expert organizations to access its text watermark detector to help evaluate the system.

OpenAI lists limitations. At a target false positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages, compared to about 95% of 400-token passages for psychology content. Detection rates are lower for subjects like mathematics where word choices are constrained. Editing also degrades the signal. In evaluations of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, while replacing 25% of words dropped detection to 17%.

The company evaluated the impact of watermarking on GPT-6 Astra, which OpenAI calls its latest frontier model, at max effort. On AutomationBench, the model scored 34.09% without watermarking and 34.86% with it. DeepSWE v1.1 scores were 72.80% unwatermarked and 71.68% watermarked. On GPQA Diamond, the scores were 94.44% without watermarking and 93.94% with it.