Google Launches First Double-Blind AI Evaluation Using Confidential Computing

Google, in collaboration with partners like MLCommons and OpenMined, has launched the world's first double-blind AI evaluation framework. This system uses Google Cloud's Confidential Space to benchmark the Gemini Flash Lite model in a secure, encrypted environment. It resolves a major dilemma in AI benchmarking where evaluators risk exposing test prompts (leading to benchmark contamination) and model developers risk exposing proprietary weights. This allows independent agencies and governments to conduct highly sensitive, unbiased AI safety tests without compromising intellectual property. The double-blind setup ensures that the evaluator cannot access the Gemini model weights, while Google cannot see the evaluator's test prompts. This cryptographic isolation prevents benchmark contamination and protects data sovereignty during high-risk external evaluations.