OpenAI is introducing textGrain, a technology designed to embed invisible statistical watermarks into text generated by its models. This move aims to comply with the European Union AI Act, which mandates machine-readable labeling for AI-generated content. The feature will be enabled for ChatGPT and Codex users in the EU in the coming weeks.
Unlike Anthropic, which mandates watermarking for its Claude models globally, OpenAI will make watermarking optional for API users worldwide. This functionality will also be accessible through cloud partners, including Microsoft Azure. The company intends to release the textGrain technology as open source to allow for broader community development.
Internal testing shows that detection rates for the watermark vary based on text length and subject matter. For 400-token passages on psychology, the system identified watermarks in approximately 95 percent of cases with a 1 percent false-positive rate. However, detection performance drops significantly when text is edited or when the subject matter involves math.
OpenAI reports that watermarking does not impact model performance, citing tests across eight benchmarks, including GPQA Diamond. The company emphasizes that the watermark does not verify authorship, accuracy, or human involvement. Access to the detection tool is currently restricted to selected researchers and organizations on a case-by-case basis.
