Сравнительный анализ поведения больших языковых моделей и людей в игре «Диктатор»

Журнал Новой экономической ассоциации. 2026. № 2(71). С. 177–203.

Shenkman E., Parshakov P., Matkin N.

This article analyzes the behavior of large language models (LLMs) in the Dictator game; the comparison was carried out against the results of previously conducted laboratory behavioral experiments with human participants. The study examines modifications of the basic game, that include the possibility of taking resources from the opponent as well as the introduction of a production phase. Ten LLMs were selected for the experiments, including both international and Russian models. The results show that LLMs systematically display greater generosity compared to human participants: the average transfers to the opponent are higher regardless of the game conditions. At the same time, the analysis of the models’ reasoning revealed the use of both fairness- and trust-related categories and arguments oriented toward maximizing self-interest, which makes their decision-making similar to the cognitive foundations observed in humans. Thus, LLMs replicate certain elements of social behavior but exhibit a shift toward “hyper-altruism,” which limits their direct applicability as substitutes for human participants in behavioral experiments.

Keywords: поведенческая экономика, social norms, социальные нормы, игра Диктатор, behavioral economics, dictator game, большие языковые модели, Large Language Models, AI Agents, ИИ-агенты