Overconfidence in People and Machines

Overconfidence is one of the most pervasive biases in human judgment. I present a theory proposing that overconfidence arises from fallibility, especially when we don’t know we are wrong. This theory predicts that any fallible reasoning agent—human or artificial—will exhibit systematic overconfidence under specifiable conditions. Testing this prediction, I examine various AI systems across different tasks, including tests of logic, reasoning, knowledge, and probability estimation. The results show that they exhibit overconfidence patterns remarkably similar to those observed in humans, including a tendency toward overconfidence, strongly moderated by task difficulty. Self-critical reflection can help improve confidence calibration. These results suggest that overconfidence stems from fallibility and error neglect rather than uniquely human cognitive limitations.

Learn more about our speaker:

Don Moore is a professor and the Lorraine Tyson Mitchell Chair in Leadership and Communication at the Haas School of Business, UC Berkeley. He received his PhD in Organization Behavior from Northwestern University. His research interests include overconfidence—including when people think they are better than they actually are, when people think they are better than others, and when they are too sure they know the truth. He is only occasionally overconfident.