Stanford researchers have uncovered major shortcomings in how AI image recognition systems are evaluated, revealing that models often fabricate visual details when no image is provided. Their new ...