OpenAI adds optional watermarking to ChatGPT and Codex in EU

The technology, called textGrain, works by embedding a statistical signal in the model’s word choices. OpenAI claims it matches or exceeds other detection methods, including Google’s SynthID, in internal tests. However, detection rates vary significantly by content length and subject matter. For example, 95% of 400-token passages about psychology were detected at a 1% false-positive rate, while math content saw substantially lower rates. Editing text—even replacing just 10% of words—reduces detection rates sharply, from about 92% to 66%. OpenAI also notes that watermarking does not affect output quality in its tests across benchmarks like GPQA Diamond or BrowseComp, though critics question whether it impacts writing quality. The company will release textGrain as open source and initially restrict detector access to researchers and specialist organizations under the EU’s Code of Practice.