Proceedings of the Annual Meeting of the Cognitive Science Society

Large language models (LLMs) are increasingly deployed in multi-agent systems involving sustained interaction and collective decision-making. While prior work has documented LLMs' individual-level social behaviors such as conformity, it remains unclear whether collective interaction gives rise to group-level psychological phenomena. Therefore, we investigate whether groupthink, a classic group-level cognitive bias, emerges in LLM-based multi-agent systems. Grounded in classical groupthink theory and human groupthink experiments, we propose GILCID framework and five quantitative metrics for investigating groupthink and its typical associated phenomena: group polarization effects, spiral of silence, and pressure for conformity. Our results show that LLM collectives exhibit human-like groupthink dynamics, leading to systematic shifts in decisions and expressed stances. We further analyze key factors shaping groupthink dynamics, including task type, group size, and group authority, and propose two training-free mitigation strategies to reduce groupthink negative effects. This work provides insights into group-level cognition in LLM collectives.