BullshitBench measures AI model robustness by presenting nonsensical prompts and scoring how models respond—either by detecting and rejecting the nonsense or incorrectly accepting it as valid. Version 2 contains 100 questions across five domains: software, finance, legal, medical, and physics. The project provides a public viewer, leaderboards tracking model performance, and analyses of factors like reasoning effort, model size, and launch date on detection accuracy.