OpenAI is rolling out its textGrain watermarking technology for ChatGPT and Codex in the EU to comply with the EU AI Act, which became effective on August 2, 2026. The system embeds machine-readable statistical patterns into text by adjusting word choice probabilities during generation, without adding visible characters or special symbols.
The rollout covers all ChatGPT subscription tiers in the EU over the coming weeks. The technology is designed to be imperceptible to humans and has minimal impact on model performance or generation speed. OpenAI notes that textGrain is not a foolproof identification method, as detection accuracy varies based on language, text length, and potential modifications.
The system does not require watermarking for code snippets or text shorter than 200 tokens (approximately 150 English words). Detection rates vary across the 24 official EU languages, ranging from 42.2% in Romanian to 69% in Spanish. OpenAI plans to open-source the technology while initially restricting access to the detection tool to approved researchers and institutions.
Unlike third-party AI detectors that analyze existing text, textGrain embeds signals during the generation process. While global API users can choose to enable this feature, it remains disabled by default. OpenAI emphasizes that this machine-readable signal does not replace mandatory visible AI labels or C2PA provenance standards.
The company warns that significant rewriting, paraphrasing, or translation of the generated content may degrade or destroy the statistical signal, rendering the watermark undetectable. Consequently, OpenAI positions textGrain as a supplementary provenance tool rather than an absolute proof of AI-generated content.
