AI-Mediated Communication Reshapes Social Structure in Opinion-Diverse Groups

TL;DR

It is found that small variations in AI-mediated communication cascade into macro-level differences in group composition, and participants with individual assistance send more messages and show greater stance-based clustering, whereas those with relational assistance use more receptive language and form more heterogeneous ties.

Abstract

Group segregation or cohesion can emerge from micro-level communication, and AI-assisted messaging may shape this process. Here, we report a preregistered online experiment (N = 557 across 60 sessions) in which participants discussed controversial political topics over multiple rounds and could freely change groups. Some participants received real-time message suggestions from a large language model (LLM), either personalized to their stance (individual assistance) or incorporating their group members' perspectives (relational assistance). We find that small variations in AI-mediated communication cascade into macro-level differences in group composition. Participants with individual assistance send more messages and show greater stance-based clustering, whereas those with relational assistance use more receptive language and form more heterogeneous ties. Hybrid expressive processes-jointly produced by humans and AI-can reshape collective organization. The patterns of structural division and cohesion depend on how AI incorporates users' interaction context.

Figures (10)

• Figure 1: Relational AI assistance reduces stance assortativity. The left panel shows deviations from baseline assortativity (Round 1), and the right panel shows estimated slopes of change with statistical comparisons against the no-assistance condition (Extended Data Table 3). Error bars represent mean $\pm$ s.e.m.
• Figure 2: Structural dynamics with and without AI assistance. Changes from the initial round are shown for (a) number of conversation partners, (b) number of conversation groups, (c) within-group stance distance, and (d) between-group stance distance across conditions. Error bars represent mean $\pm$ s.e.m.
• Figure 3: Relational assistance increased message receptiveness, while individual assistance sustained communication volume. Individual-level message counts (a) and sentiment (b), shown with and without AI-assisted messages. Group-level changes in message counts (c) and sentiment (d) across rounds. Error bars represent mean $\pm$ s.e.m.; $p$-values from regression analyses.
• Figure 4: Heavier AI users sent more messages. Dots represent individual participants’ AI suggestion use rate and their message counts per round under individual and relational assistance conditions ($N=182$ in the individual assistance condition; $N=188$ in the relational assistance condition). $r$ indicates Spearman’s rank correlation; both correlations are significant at $p < 0.001$.
• Figure 5: Effects of message sentiment and volume on ego-centric stance distance. (a) Example of changes in ego-centric stance distance across rounds. From Round $t$ to Round $t{+}1$, one member left and two members joined the group (bold edges). Node color indicates each member's stance (see Fig. \ref{['fig:networks']}). (b) Predicted between-round change in ego-centric stance distance for participants who remained in the same group, estimated using a linear mixed model (Extended Data Table 6). The contour line marks the transition point ($\Delta = 0$) between reduced and increased stance distance. Axes represent average sentiment (VADER score) and number of messages sent. Marginal density plots show the observed distributions of message sentiment and message counts across the three AI assistance conditions.