OpenAI will add visible watermarks to ChatGPT text in the EU; API users can globally opt out

To comply with the EU AI Act, OpenAI will enable invisible text watermarking for ChatGPT and Codex users in the EU in the coming weeks. The technology, named textGrain, works by embedding statistical signals in the model's word choices. Unlike Anthropic's mandatory watermarking for Claude globally, OpenAI's API watermarking is opt-in globally and will be available in the coming weeks through cloud partners such as Microsoft Azure. OpenAI says that in internal testing, textGrain matches or exceeds approaches like Google SynthID, and it plans to open-source the technology. Detection rates depend heavily on text length and topic: at a 1% false positive rate, detection for a 400-token psychology passage is about 95%, dropping to about 80% at 200 tokens, and significantly lower for mathematical content. Editing substantially weakens detection: replacing 10% of synonyms reduces detection for a 400-token passage from about 92% to 66%, and replacing a quarter drops it to 17%.