A large-scale study tested whether AI personas can detect when humans are lying—and found that while AI can sometimes spot deception, it’s still far from trustworthy.