DeepMind pilots double-blind evaluations of Gemini

Google DeepMind said on August 27 it is running what it calls the first double-blind evaluation of a proprietary frontier-class model, testing a Gemini Flash Lite system against confidential benchmarks inside a cryptographically sealed Google Cloud Confidential Space environment. Partners include the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons; the setup is designed so evaluators cannot see Gemini weights and Google cannot see the test prompts, addressing benchmark contamination as models get more capable.