Problem

The increasing integration of artificial intelligence (AI) agents into human group settings raises critical questions about their impact on cooperative social norms. While previous research has largely focused on dyadic (two-person) human-AI interactions, often observing a "machine penalty" where cooperation with AI is lower, there is limited understanding of how AI agents affect the emergence and maintenance of cooperation within small groups. This study aims to address the theoretical tension between two potential outcomes: "differentiation," where the artificial identity of an agent disrupts social cohesion, and "normative equivalence," where the functional mechanisms of group norms render the agent's identity irrelevant, thus questioning whether the presence of an AI fundamentally alters the dynamics of group cooperation.

Method

To investigate these dynamics, an online experiment was conducted using a repeated four-player Public Goods Game (PGG), followed by a one-shot Prisoner's Dilemma (PD) game. Each group consisted of three human participants and one computer-controlled bot. The experiment employed a 2x3 factorial design, manipulating two key variables: the agent label (the bot was presented as either "human" or "AI") and the bot's predefined decision strategy (unconditional cooperator, conditional cooperator, or free-rider). Participants were unaware of the bot's strategy. The study measured participants' contributions in the PGG over ten rounds, their cooperation decisions in the subsequent PD, and their normative perceptions, including social appropriateness ratings and empirical and injunctive norm expectations. A sample of 236 participants was recruited via Prolific. Data analysis involved linear mixed-effects regressions for the PGG, logistic regression for the PD, and examination of norm perceptions.

Results

The study found no statistically significant difference in cooperation levels in the Public Goods Game between groups with a human-labelled bot and those with an AI-labelled bot, leading to the rejection of the initial hypothesis of a "differentiation effect." Instead, cooperation was primarily driven by reciprocal group dynamics (responsiveness to others' prior contributions) and behavioral inertia (one's own prior contributions), alongside a typical gradual decline over time. The specific strategy of the bot (unconditional cooperator, conditional cooperator, or free-rider) also had negligible impact, suggesting that the presence of other human group members buffered against extreme bot behaviors. Furthermore, in the one-shot Prisoner's Dilemma, there were no systematic differences in norm persistence based on the bot's label or strategy. Participants' normative perceptions, including social acceptability ratings and both empirical and injunctive expectations, were remarkably consistent across human and AI conditions. While trust was generally a significant predictor of higher contributions, and normative pressure played a role in AI groups, a general "AI acceptance" in AI groups was surprisingly associated with lower contributions.

Implications

These findings introduce the concept of "normative equivalence," demonstrating that the fundamental mechanisms sustaining cooperation, such as reciprocity and conditional cooperation, operate similarly in both mixed human-AI and all-human groups. This challenges the common assumption of algorithm aversion in group contexts, particularly when humans constitute the majority. The results suggest that in minimal social presence and communication scenarios, individuals prioritize observable group behavior and social cues over the ontological identity of an agent (human versus AI). Practically, this implies that designing AI systems for cooperative teams might benefit more from focusing on transparent, norm-consistent behavior than on anthropomorphic design or human-like labels. However, the study acknowledges limitations, including the possibility of participant disbelief in group composition, the short-term and anonymous nature of interactions, the use of scripted rather than adaptive AI, and the aggregate feedback mechanism potentially diluting the bot's individual strategy. Future research could explore adaptive AI agents, varying proportions of AI in groups, and transparent individual feedback to understand when and how this observed normative equivalence might break down.