New AI Benchmark Tests Whether Models Know When to Question Their Own Decisions

A developer built a benchmark called Decision-Conditioned Falsification to test whether AI models can rationally decide when gathering more information is worth the cost before acting. The environment presents three possible hidden mechanisms, and models must compose experiments and commit to operational actions, with experiments costing points and poor decisions costing even more. Unlike typical benchmarks, it rewards net operational utility rather than the volume of self-criticism, hypothesis revisions, or experiments run. Testing across four models — GPT-4.5 nano, Claude Haiku 4.5, Gemini 3.5 Flash, and Gemini 3.7 Flash — revealed no consistent aggregate benefit from explicitly prompting models to seek contradicting evidence. A key finding was that models sometimes predicted outcomes impossible under any mechanism, suggesting the core problem is not willingness to self-correct but the ability to accurately calculate what a hypothesis actually predicts.