OpenAI said it will add invisible watermarks to text generated by ChatGPT and Codex for users in the European Union, rolling out to eligible accounts on all plans over the coming weeks, according to TechCrunch. The feature, called textGrain, subtly shapes the model's word choices to embed a statistical pattern that a detector holding a secret key can identify, though a human reader cannot see it, TechCrunch reported.
The rollout responds to the EU AI Act's transparency rules, which took effect Aug. 2 and require AI companies to mark AI-generated content in identifiable ways, according to TechCrunch. API developers anywhere in the world can turn the watermark on starting immediately, though it stays off by default outside the EU, TechCrunch reported.
OpenAI said "the watermark doesn't identify the user" and that it saw no meaningful change in model performance with it switched on, according to TechCrunch, which reported that OpenAI published a technical report on the system with researchers from the University of Pennsylvania and Yale.
The watermark has limits. TechCrunch reported that replacing just 10% of a text's words with synonyms cut detection from 92% to 66% in OpenAI's own testing, and the company cautioned that "a missing watermark does not prove human authorship."
A watermark that falls apart after a light rewrite will not stop anyone determined to pass AI text off as human, but it does give platforms and regulators a default signal to check first. For developers building on OpenAI's API, the compliance line is now explicit: inside the EU, mark it, or face the AI Act's transparency rules directly.