Google Launches World's First Double-Blind AI Benchmark Test

Google has introduced the world's first double-blind evaluation of a frontier AI model, using cryptographic technology to prevent benchmark contamination. The pilot tests a Gemini Flash Lite model against confidential benchmarks where neither Google nor evaluators can access each other's proprietary data. Partnering with Singapore's AI Safety Institute, OpenMined, AVERI, and MLCommons, Google used Confidential Space within Google Cloud to ensure model weights and test prompts remain private to their respective owners. The approach aims to build greater trust in AI benchmarks among policymakers, researchers, and enterprises.