A study found the same AI models reached 94.9% accuracy on medical scenarios when experts ran them, dropping to 34.5% when the public did. We've been seeing the same pattern in our own work.