Why self-driving AI fails when cameras fog up

Self-driving systems that explain their reasoning (like saying "turning left because pedestrian detected") break down unpredictably under real-world sensor noise, fog, and lighting extremes. Researchers stress-tested Alpamayo, a 10B-parameter driving model, across 1,996 scenarios with eight types of sensor corruption and found that when the AI's explanations change after perturbation, trajectory errors spike dramatically. The key finding: explanation consistency is a 0.99-correlation proxy for actual safety—if the reasoning shifts, the car will deviate from its path. This suggests AI systems could use their own explanations as a built-in safety monitor during deployment.