Piloting the world's first double-blind AI evaluations

Double-blind AI evaluations are a new testing method where external evaluators assess AI models without the model provider seeing the test questions and without evaluators seeing the model's internal weights, using cryptographic safeguards to keep both sides private. This approach matters because AI models can artificially inflate their performance scores if they have already encountered test questions during training, a problem called benchmark contamination that undermines the trustworthiness of evaluation results. Historically, organizations faced a difficult choice: share test prompts and risk the model provider seeing them in advance, or share the model itself and expose proprietary information. By using cryptographic verification technology, double-blind evaluations eliminate this compromise and allow independent organizations to rigorously test advanced models while protecting both intellectual property and data security.