Гендерные различия в игре диктатора: сравнение поведения больших языковых моделей и людей

Вестник Пермского университета. Серия: Экономика. 2026. Т. 21. № 1. С. 42–57.

Parshakov P., Paklina S., Matkin N., Shenkman E.

Introduction. Large language Models (LLM) are increasingly being used in social sciences to simulate the behavior of experimental participants and analyze norms of cooperation and justice. However, the question remains whether they are capable of reproducing social asymmetries, including gender differences. Goal. The work aims to test whether LLM reproduces gender differences in the Dictator game and how textual rationales for decisions relate to the chosen strategies. Materials and methods. The classic experiment of Eckel and Grossman (1998) is adapted for analysis, where the "dictator" distributes 10 dollars between himself and his opponent. Five models (GPT-5, Grok-4, YandexGPT 5 Pro, Gigachat 2 Max, and Qwen3-235b-a22b) generated 1,000 decisions based on twenty participant profiles. The comparison was carried out with the results of experiments with human subjects; additionally, regression analysis was applied, taking into account textual explanations. Results. All models demonstrated higher generosity compared to humans: "male" agents transferred an average of 3.57 units, "female" — 4.18, while human subjects — 0.82 and 1.6, respectively. The gender gap persists, but decreases with the control of textual arguments. Mentioning "equality" and "justice" is associated with an increase in generosity, while appeals to "rationality" and an emphasis on "the opponent" are associated with a decrease in generosity. Significant differences were also found between the models: Grok-4 and YandexGPT 5 Pro had the most generous behavior, while Gigachat 2 Max and Qwen3-235b-a22b had the least. Conclusions. LLMs reproduce both norms of cooperation and social asymmetries, while their severity depends on architecture and argumentation strategies. The results demonstrate that the behavioral assessment of LLM should be combined with the analysis of explanations of decisions. The study is aimed at researchers in the field of behavioral economics, AI, and social informatics; the prospect is to expand the analysis to n-player games, multilingual protocols, and control of inference parameters.

Keywords: поведенческая экономика, искусственный интеллект, экономические агенты, экспериментальные исследования, экономические решения, economic agents, моделирование поведения, behavior modeling, experimental research, social interactions, социальные взаимодействия, behavioral economics, большие языковые модели, artificial intelligence, Large Language Models, synthetic samples, economic decisions, синтетические выборки