.png){kind=icon}
.png){kind=icon}
Return
Prosocial When Simple and Cold-Hearted When Complex: How Task Difficulty Shapes LLM Behavior
.png){kind=icon}
PRE
AI
DOI：10.1287/deca.2025.0396.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Prior studies suggest that large language models (LLMs) act prosocially in simplified game-theoretic settings, but whether such behavior reflects stable objectives or context-driven patterns is unclear. We test whether LLMs exhibit fairness when choices follow complex tasks or take place in more complex decision environments. We hypothesize that problem complexity and mathematical prompts increase the LLM's weight on prioritizing self-interests by activating responses geared toward calculation and rationality. We operationalize our theory using a quantal response framework and conducted a series of experiments with GPT-4, GPT-4o, and o3-mini as decision makers to test our hypotheses. In Study 1, models played Dictator and Ultimatum games following a series of unrelated problems that varied in context and difficulty. Study 2 was a sequential supply chain game that mirrors key aspects of the Ultimatum game regarding fairness concerns, but with added complexity. In Study 1, simple prompts produced nearly equal splits, because of fairness norms and preference for equity. However, complex math prompts invoked rational profit maximization logic to reduce allocation offers. In the pricing game, the models prioritized self-interested pricing but differed in decision execution. GPT-4 and GPT-4o selected lower prices because of random errors and heuristic responses rather than fairness concerns. In contrast, o3-mini consistently derived the profit-maximizing solution. Fairness in LLM responses is context sensitive and often suppressed by task characteristics that trigger goal-directed responses. Thus, researchers and developers must assess social preferences in more complex scenarios. Moreover, our research shows that utility-based models that incorporate bounded rationality and fairness capture core patterns in LLM behavior and yield testable predictions, supported by both choice data and model-generated text.
Keywords:
large language models
GPT
fairness
task complexity