Overconfident AI: how artificial intelligence navigates risk and uncertainty
Sounman Hong, Sanghyun Lee · Journal of Psychology and AI · 2025
To investigate differences between artificial intelligence (AI) and human decision-making, we developed an AI system to play Go, an ancient board game that demands complex decision-making in uncertain environments. While AI systems vary widely in design and decision-making strategies, this study focuses on a specific AI trained through self-play. By analysing games between this AI and human players of comparable abilities, our study presents two main findings. First, our AI system is more prone to making severe, game-costing errors compared to their human counterparts. We attribute this behaviour partly to differing risk preferences; while humans display risk-averse tendencies, the AI system exhibited risk-neutral behaviour. Specifically, AI systems tend to make decisions associated with higher expected returns but also with higher risks compared to humans; these risky decisions sometimes result in catastrophic errors. Our findings suggest that AI systems, when designed without adequate mechanisms to account for risk, may exhibit overconfidence in uncertain environments.