OpenAI to Watermark ChatGPT Text in the EU, Limits Detector Access

OpenAI said on Monday it will begin embedding an invisible watermark in text generated by ChatGPT and Codex for eligible users in the European Union, rolling it out over the coming weeks across all plans. The company is not making watermarking a global default.

The move responds to the EU AI Act's transparency rules, which took effect on August 2 and require AI companies to mark generated content in a way other systems can detect.

For developers, the change is optional. Starting today, API customers anywhere can turn on watermarking for select models, though it remains off by default. OpenAI said it is also working with cloud partners to offer watermarking for its models through their platforms in the coming weeks.

Checking for the watermark will be harder. OpenAI opened applications for its text detector but is limiting access to approved researchers and expert organizations, citing the risk of false positives and missed watermarks. The detector reports only whether an OpenAI watermark is present, without identifying the user or revealing prompts. OpenAI's image and audio verification tools remain public.

The method, called textGrain, places a hidden statistical pattern in the model's word choices. OpenAI published a technical report co-written with researchers from the University of Pennsylvania and Yale, and says it will release the technology as open source.

OpenAI's own results show clear limits. At a 1% false positive rate, its detector caught watermarks in about 80% of 200-token passages and about 95% of 400-token passages on psychology questions, and did substantially worse on math. Swapping 10% of the words for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%.

"Strong performance under ideal conditions does not guarantee reliable detection in everyday use," the company wrote.

OpenAI also claims textGrain matched or beat other methods it tested, including Google's SynthID, and that watermarking caused no meaningful drop in benchmark scores for its Astra model. Both claims are self-reported; AI Briefing News has not independently verified them.

OpenAI did not say which API models support watermarking, which ChatGPT outputs count as eligible, how it will approve detector applicants, or when the detector might open to the public.

OpenAI's approach is narrower than Anthropic's. Anthropic began marking text from Claude models released after August 2 and is applying the watermark globally rather than only in Europe, according to The Next Web. Its detection API is in private preview and also available to enterprises that need to verify watermarks for their own compliance.