OpenAI announced that it will introduce invisible watermarks for text generated by ChatGPT and Codex in the European Union, aligning with the EU AI Act transparency rules that took effect on August 2. The measure uses a technology called textGrain to embed a machine-readable signal in the model's word choices. The company described the rollout as phased and limited to the EU at launch while making opt-in access available globally to API customers.
Compliance Scope and Rollout Timeline
The watermarks will appear over the coming weeks for eligible ChatGPT and Codex users across all plans in the EU only. OpenAI said it is not making text watermarking a global default at launch to allow time for real-world feedback. Starting today, developers using OpenAI's API anywhere in the world can activate watermarked outputs for select models, although the feature remains off by default.
How textGrain Embeds an Invisible Signal
OpenAI's textGrain method is not a visible symbol but a subtle shaping of the model's word choices, creating a pattern that readers cannot see and a detector can identify. Because the signal is embedded in the language itself, it remains present when text is copied and pasted. The watermark does not identify the user, and OpenAI reported no meaningful change in model performance with the feature switched on.
Detection Reliability and Known Limits
OpenAI co-published a technical report on textGrain with researchers from the University of Pennsylvania and Yale. The evaluation shows detection is strongest for longer passages and weaker for short, mathematical, or translated content. In one test, replacing 10% of words with synonyms reduced detection from about 92% to 66%, and replacing 25% cut it to 17%.
Researcher Access and Responsible Use
The company will initially provide detector access only to approved researchers and expert organizations to support evaluation and improvement. This decision reflects the risk of false positives and missed watermarks. OpenAI also cautioned that the absence of a watermark does not prove human authorship because text may be too short, heavily edited, translated, or generated by another provider.
What a Watermark Does Not Prove
OpenAI stressed that a text watermark does not measure human contribution, establish ownership, or determine responsibility for a passage. It also does not verify accuracy, tell whether content is misleading, or identify the user. A detected watermark can only indicate that an OpenAI system generated or processed part of a passage.
Industry Context and Market Response
The announcement follows Anthropic's decision two months ago to watermark Claude outputs worldwide, which drew backlash from some users who argued the model merely executed their instructions. OpenAI had previously built a text watermark but delayed release partly over concerns that users would switch to competitors without watermarking. OpenAI, Anthropic, Google, Meta, and Microsoft have all committed to the EU's code of practice on AI-generated content.
OpenAI's EU rollout represents a cautious step toward machine-readable text provenance, balancing regulatory compliance against known technical limits. By keeping the watermark regional for ChatGPT and Codex, while allowing global API customers to opt in, the company is testing adoption and reliability before any wider expansion. The company signaled it will revisit the approach as detection methods, standards, and evidence evolve.