OpenAI released text watermarking for select models in the EU on October 5, 2026, to comply with the EU AI Act. It is the company's first safety update since its initial content provenance tools.
OpenAI reported textGrain, its watermarking technology, matched or exceeded other approaches like SynthID for text, measured on benchmark tests. That compares with previous methods, which had lower detection rates.
Text watermarking is built on statistical signals added to word choices and targets transparency for AI-generated text. Availability begins with API customers globally, initially for select models.
"Our phased approach reflects both the EU AI Act requirements as well as the technology’s limitations," said OpenAI. The company emphasized transparency about what a text watermark can and cannot tell people.
The announcement follows the EU AI Act's requirement for generative AI providers to make generated text identifiable. OpenAI framed the move as part of its broader work on content provenance.
OpenAI did not say how widely the watermarking will be used beyond the EU, and raised concerns about false positives and negatives. The company expects to revisit the approach as the technology evolves.
Source: openai
