Illusions of Confidence in Artificial Systems

Effective collaboration requires that we monitor both the cognitive states (e.g., beliefs) and metacognitive states (e.g., confidence) of other agents. While humans routinely share confidence, metacognitive capabilities are still developing in artificial intelligence (AI), raising the question of how humans attribute metacognition to AI systems. In seven pre-registered experiments, we show that attributions of metacognition are sensitive to observed behaviour (e.g., response times), but also agent types: observers consistently overestimated AI confidence compared to humans—even when their behaviour was identical. This illusion of confidence was robust across behavioural profiles, agent descriptions, and decision-making tasks (visual perception, general knowledge) but was reduced in more subjective decisions (emotion categorisation). An experimental manipulation further showed that illusions of confidence are rooted in prior beliefs about the agents’ capabilities. Together, these findings uncover a powerful illusion of confidence in artificial systems and highlight a central role for metacognition in human-AI interactions.