Do Large Language Models Resolve Fairness-Efficiency Trade-offs Like People?
Abstract
With rapid improvements in reasoning, large language models (LLMs) have gained more prominent roles as decision-makers in social and organizational settings. This makes it important to understand whether they adhere to the same values as people about division of labor. Here, we test whether LLMs and humans are aligned on task allocation problems. These problems present a trade-off between efficiency (optimizing for metrics like output and completion time) and fairness (where each collaborator must complete roughly an equal or equitable share of the overall workload). In two experiments we find that people and LLMs vary in how closely their stated preferences between efficiency and fairness align. Humans and LLMs frequently diverged from their stated preferences and settled on similar allocations when actively determining an allocation on their own. Our work identifies a value-action gap in both humans and LLMs that influences the degree to which LLMs align with social preferences.