A new psychology study reveals that when a highly accurate AI system confidently accuses someone of lying, human supervisors tend to override the decision—a "truth bias" that actively destroys the AI's detection accuracy.