Noncooperative Human-AI Agent Dynamics

A team of researchers including Dylan Waldner has published a comprehensive study titled 'Noncooperative Human-AI Agent Dynamics' on arXiv, exploring strategic competition between artificial intelligence and human decision-makers. The 41-page paper introduces a novel modeling framework where AI agents operate as standard expected utility maximizers, while human agents are more accurately represented using Prospect Theory from behavioral economics. This approach incorporates well-documented cognitive heuristics like reference dependence and loss aversion—where humans feel losses more acutely than equivalent gains—creating a more realistic simulation of human strategic behavior.

The researchers conducted extensive numerical simulations across various classic matrix games and specialized scenarios designed to highlight differences in strategic approaches. They tested interactions between three agent types: pure AI agents, 'aware' humans with full game knowledge, and learning Prospect Agents that simulate human-like decision-making. The results revealed a spectrum of emergent behaviors, from situations where AI and human agents were barely distinguishable to patterns that confirmed known Prospect Theory anomalies, along with some unexpected strategic surprises. The study provides both code and detailed analysis of how these mixed-population competitions unfold, offering new insights into human-AI interaction dynamics in competitive environments.