OpenAI Deploys Invisible Watermarking for ChatGPT in the EU

The implementation marks a shift in how OpenAI manages its transparency obligations, focusing on a regional rollout rather than a global default. While the company claims textGrain performs on par with Google DeepMind’s SynthID, it emphasizes that these watermarks cannot prove human authorship, measure human contribution, or confirm the ownership of the generated text. Internal benchmarks suggest that applying these markers does not degrade the performance of the models.

Simultaneously, OpenAI is offering an opt-in version of this technology for API customers worldwide. This allows developers to integrate watermarking based on their specific transparency needs. Access to detection tools is also being extended to approved researchers and expert organizations, which can begin applying for access immediately. Further integration is planned through cloud partners to ensure consistent coverage across various third-party services in the coming weeks.