Human–AI social cooperation: convergences, divergences, and challenges

Cooperation, a cornerstone of social interaction, not only shapes individuals’ behavioral decisions but also profoundly influences group cohesion and social evolution. From both evolutionary and cultural perspectives, humans demonstrate an “ultra-social” nature, characterized by consistent cooperation behaviors such as food sharing, caregiving, and mutual support observed across a wide range of ecological and societal contexts. From small foraging societies to large nations, individuals consistently demonstrate a strong willingness to cooperate with others. This cooperative capacity starts showing as early as middle childhood, highlighting deep evolutionary roots and critical adaptive advantages. The rapid development of artificial intelligence (AI), particularly large language models (LLMs), has introduced a new generation of AI systems with remarkable capabilities in language comprehension, social reasoning, and contextual judgment, enabling them to participate in conversations, infer intentions, and respond to nuanced social cues in “human-like” ways. As AI becomes increasingly integrated into diverse social contexts and takes on more and more complex, varied roles, human–machine interactions are undergoing a profound transformation. This shift challenges our traditional theoretical frameworks of social interaction and raises urgent questions about the underlying mechanisms of cooperation when one or more partners are artificial agents.To address this trend, this review systematically examines classic cooperation research and identifies three core stages of the cooperative process: strategy formation, behavior execution, and outcome evaluation and learning. Each stage encapsulates distinct cognitive operations and interaction dynamics. This continuous framework highlights that humans and AI not only exhibit distinct cognitive characteristics and behavioral modalities at each stage, but also manifest structural distinctions in inter-stage transitions. Within this framework, we summarize major findings from human interpersonal cooperation and human–AI cooperation. In addition to this, we compare significant differences in decision-making, cooperation patterns, and feedback adaptation between these two types of interaction. For example, although LLMs increasingly exhibit “human-like” cooperation tendencies, sometimes even outperforming human participants, these behaviors are driven by fundamentally different underlying processes. Moreover, when engaging with AI partners, humans often adjust their expectations differently than with human partners. This can lead to changes in trust formation, a slower development of shared routines, and an increased tendency to interpret AI behaviors through anthropomorphic lenses. Such patterns indicate cognitive and social adaptations unique to human-machine settings.Based on these comparisons, we propose several strategies for optimizing AI systems in cooperative contexts, including enhancing the interpretability of AI decisions, designing more flexible real-time feedback mechanisms that respond to users, and ensuring transparent boundaries of agency to support smoother role negotiation between human and artificial agents. These insights enrich the current discourse on cooperation among both biological and artificial agents, providing practical guidance for advancing human-AI interaction models. By focusing on these areas, this review aims to deepen the understanding of cooperation between biological and artificial agents, and in doing so, it offers concrete suggestions for improving upcoming human-AI interaction models. It also underlines the crucial role of cognitive science and psychology in developing ethical, transparent, and socially responsible forms of cooperation in an increasingly blended human-AI society.