Proceedings of the Annual Meeting of the Cognitive Science Society

Human decision-making is filled with a variety of paradoxes demonstrating deviations from rationality principles. Do state-of-the-art artificial intelligence (AI) models also manifest these paradoxes when making decisions? As a case study, in this work we investigate whether GPT-4, a recently released state-of-the-art language model, would show two well-known paradoxes in human decision-making: the Allais paradox and the Ellsberg paradox. We demonstrate that GPT-4 succeeds in the two variants of the Allais paradox (the common-consequence effect and the common-ratio effect) but fails in the case of the Ellsberg paradox. We also show that providing GPT-4 with high-level normative principles allows it to succeed in the Ellsberg paradox, thus elevating GPT-4's decision-making rationality. We discuss the implications of our work for AI rationality enhancement and AI-assisted decision-making.