Abstract
<p>Using a corpus of 51 mock-jury discussion groups, we analyzed the discussion process of human groups and then compared it with that of large language model (LLM) groups in an exploratory analysis. The task was a hidden profile task, solvable only by combining information distributed across members. We examined solution-path attainment and the sharedness of the word-level semantic network (the cross-member edge ratio: the proportion of edges linking words introduced by different members) in relation to the information-sharing process. Human discussion progressed from individual to group-level sharing: an information-sharing phase in which intra-individual semantic networks remained (cross-member edge ratio .34), followed by an information-integration phase in which network links became shared across members (.54), and a decision phase (.66). In LLM groups calibrated to human behavioral statistics, this three-phase structure was radically compressed, lacking the human allocation of time. Under a memory constraint, however, the individual-to-group sharing shift observed in humans emerged. A stepwise prompt ladder reproduced the reported collapse of LLM collective reasoning and identified the constraints that avert it. Moreover, the association between the sharedness of the semantic network and solution-path attainment found in human groups (ρ = +.52, n = 50) was not observed in LLM groups (n = 8 per condition). These differences are interpreted as showing the contrast between human groups, which come to share a semantic network gradually through discussion, and LLM groups, which share a knowledge structure from the outset.</p>