Artificial intelligence is no longer just processing language—it’s making decisions. A new study led by HKUST’s Professor Songfa Zhong and his colleagues examines whether large language models (LLMs) like GPT exhibit economic rationality, a fundamental principle in decision-making. The findings? GPT’s choices are not only rational but, in many cases, more consistent than those of humans.
The study explores how GPT makes decisions in four economic domains: risk, time, social, and food preferences. Economic rationality refers to how consistently a decision-maker allocates resources to maximize benefits while respecting constraints like budgets or risks. It is assessed using the Generalized Axiom of Revealed Preference (GARP), a standard economic test that evaluates whether choices align with logical decision-making .
To test its decision-making, researchers assigned GPT 10,000 economic tasks, where it had to distribute resources under different conditions—such as choosing between risky investments, immediate versus delayed rewards, self-interest versus social benefit, and food preferences (meat vs. vegetables). The AI’s performance was compared to 347 human participants making the same choices .
The results were striking. Think of rationality scores like a “decision-making accuracy meter,” where 1.000 represents perfect logical consistency. GPT scored nearly perfect across all four domains—99.8% in risk, 99.7% in time and social decisions, and 99.9% in food choices. Humans, while still highly rational, scored slightly lower—98.0% in risk, 98.5% in time, 96.7% in social, and 96.3% in food decisions. While these differences may seem small, they are statistically significant, meaning GPT’s decision-making was measurably more consistent than that of human participants .
Interestingly, GPT’s decision-making was highly stable across different demographic conditions, such as age and gender, while human rationality varied based on these factors. However, the AI’s rationality dropped significantly when decision questions were framed differently or when forced to choose from a set of predefined options rather than having continuous flexibility .
What does this mean for the future? The study suggests that LLMs like GPT could assist in financial advising, economic modelling, and policy-making, given their ability to make consistent, rational decisions. However, their sensitivity to question framing raises concerns about potential biases in real-world applications.
As AI continues to evolve, understanding its decision-making strengths and limitations will be crucial. If GPT can outperform humans in economic rationality, could it soon be trusted to help us make critical financial and policy decisions? The debate is just beginning.