OpenAI rolls out text watermarking for EU users and API opt-ins

The company’s technology, named textGrain, adds an invisible statistical signal to word choices. OpenAI states that in its evaluations, textGrain matched or exceeded the performance of other approaches, including SynthID for text. However, the company acknowledges significant limitations: detection rates drop for shorter or constrained text, and editing can weaken the watermark. For example, replacing 25% of words with synonyms reduced detection rates from about 92% to 17% in 400-token passages.

Access to the text watermark detector is initially limited to approved researchers and expert organizations to help evaluate reliability and responsible uses. OpenAI emphasizes that a watermark does not establish ownership, identify the user, or verify accuracy. The company plans to make the technology open source and will continue to refine its approach as standards and evidence evolve.