Outcome Without Mechanism: LLMs Reproduce Group Polarization by Imitation, Not Social Comparison

Large language model (LLM) agents in group discussion reliably produce group polarization: post discussion opinions grow more extreme in the group’s dominant direction. Prior work establishes this outcome but not its mechanism. In humans, polarization is driven by two dissociable processes Persuasive Arguments Theory (PAT, informational influence from novel arguments) and Social Comparison Theory (SCT, normative influence from learning others’ positions) which respond in opposite ways to two moderators: response privacy and task type. We port the classic PAT/SCT dissociation to LLMs with a factorial design that turns each mechanism on and off (argumentswithout-positions vs. positions-without-arguments), crossed with private vs. public response framing and intellective (Choice Dilemmas) vs. judgmental (attitude) items, across 6 models via OpenRouter with 40 seeds per cell (21,120 trials). LLMs robustly reproduce the polarization outcome. The argument mechanism behaves as in humans: novel arguments shift opinions strongly (d= 1.03 intellective, 1.51 judgmental) and equally whether the response is private or public. The positions mechanism does not: exposure to peers’ bare numbers shifts opinions substantially even under fully private, anonymous response; it does not overshoot the group (no one-upmanship); and it is strongest on intellective items, reversing the human task-type moderator. This is the signature of imitation, not audience-gated social comparison and its strength varies sharply across models. LLMs can thus match a population-level social outcome while running a non-human causal process, which bears directly on their validity as social simulators.