Research Reveals Unpredictable LLM Behavior in Economic Decision-Making: Study Finds Models Exhibit Altruism, Risk Aversion in Ultimatum Game
Key Takeaways
- ▸LLM behavior in strategic economic settings is heterogeneous but predictable when conditioning on stake size and opponent type
- ▸Different models exhibit distinct behavioral modes: some rational, some mimicking human social preferences, and some displaying unexpected altruism with distributions exceeding 50%
- ▸LLM proposers forgo significant payoffs—even more so when playing against humans—suggesting risk-averse and potentially inefficient economic decision-making
Summary
A new working paper from the National Bureau of Economic Research examines how large language models behave in the Ultimatum Game, a classic economics experiment testing fairness and rational decision-making. Researchers Douglas K.G. Araujo and Harald Uhlig tested various LLMs across different stake sizes and opponent types (human vs. AI), finding that model behavior is heterogeneous but predictable based on contextual factors. The study reveals three surprising patterns: LLM behavior varies significantly across models and conditions; some models approximate rational economic theory while others exhibit human-like social preferences or an "altruistic mode" proposing hyper-fair distributions exceeding 50%; and proposers consistently forgo substantial payoffs, especially when playing against human opponents. The findings raise important questions about deploying autonomous AI agents in economic and financial settings without thorough behavioral testing.
- The research highlights critical need for behavioral testing before deploying AI agents in real-world economic and financial applications
Editorial Opinion
This research underscores a fundamental challenge in AI deployment: models trained on human language and knowledge may inherit human biases and social preferences without explicit instruction to do so. While the emergence of fairness-seeking behavior might seem positive from a societal perspective, the inconsistency and unpredictability across models and contexts raises serious concerns for financial and economic applications where rational decision-making is paramount. The finding that LLMs forgo significant payoffs suggests they may be unreliable agents in zero-sum or competitive scenarios, warranting careful evaluation before integrating them into autonomous trading, allocation, or negotiation systems.
