Consensus-building conversation leads to neural alignment

The study investigates how consensus-building conversation changes private beliefs and brain responses over time. While prior research has established that conversation can shape attitudes and behaviors and that these effects can spread through social networks, it is unclear whether such conversations change how individuals fundamentally perceive and interpret the world. The authors distinguish private acceptance from public compliance and propose that neural alignment (measured via inter-subject correlation, ISC, of fMRI BOLD signals) can index private acceptance. Building on work showing that shared interpretations and memories correspond to shared neural responses, the central hypothesis is that group discussion aimed at consensus will increase neural alignment among group members and that this alignment will generalize to novel but related stimuli. The study also examines how real-world social network position (centrality) and perceived social status relate to conversational dynamics and neural alignment.

Prior work shows that shared context or instructions can align neural responses across individuals. Studies have found that when participants adopt different roles (e.g., detective vs. interior decorator) or receive different interpretations of an ambiguous story, their brain activity aligns within similarly instructed groups and diverges between groups. Alignment has been linked to engagement and learning in classrooms and is higher among friends within real-world social networks, suggesting social relationships relate to neural similarity. Research on social networks indicates interactivity predicts centrality, implying that alignment within networks may depend on individuals’ social behaviors. However, many prior alignment studies used top-down instruction or isolated interpretation, whereas real-world conversation is interactive and reciprocal. This study addresses that gap by examining brain alignment produced by naturalistic, consensus-seeking conversations and by relating alignment to conversation behavior, perceived status, and social network centrality.

Participants: First-year MBA students at a private U.S. university. n = 49 (23 male, 26 female by self-report; age 26–32, mean ~27.6) completed the fMRI study; 9 additional participants formed a no-conversation control group. Groups: 9 conversation groups (mean group size 4.2). All participants had social network data from a cohort-wide survey to compute network centrality (eigenvector centrality and brokerage) and a PCA composite (PCA centrality). Ethics: Informed consent; Dartmouth IRB approval. Design: Three sessions. Session 1 (pre-conversation): fMRI while viewing ambiguous movie clips (no sound) designed to support multiple plausible interpretations; participants then completed a survey about each clip’s narrative. Session 2 (conversation): small groups discussed the clips (15 min/clip) to reach consensus; they then individually rated their agreement with the group consensus and reported perceived influence of group members. Session 3 (post-conversation): fMRI while viewing the same clips plus novel clips from later segments of the same movies, followed by surveys about the novel clips. The control group completed both fMRI sessions without intervening conversation. Imaging acquisition and preprocessing: EPI images motion-corrected (FSL mcflirt), motion outliers flagged (framewise displacement > 0.9), EPI–anatomical alignment (AFNI align_epi_anat.py), percent scaling and MNI normalization via concatenated transforms (ANTs antsApplyTransforms), spatial blurring (AFNI 3dBlurInMask), and nuisance regression (motion, outlier regressors, tissue counts, linear/quadratic trends, intercept). Hyperalignment was performed with PyMVPA using naturalistic movie runs with sound to improve voxelwise correspondence across participants. Voxelwise inter-subject correlation (ISC) was computed as Pearson correlation of time series at corresponding voxels between participant pairs (AFNI 3dCorrelation). Behavioral measures: Surveys assessed individual and group beliefs about narratives. Behavioral similarity/distance between participants’ survey responses was modeled with hierarchical linear regression including session (pre vs post) and comparison type (within vs between group) with random intercepts for participant pairs. Neuroimaging analyses: Conversation-induced change in ISC was quantified by subtracting pre- from post-conversation ISC for each participant pair, yielding a change matrix. Multiple regression modeled change in ISC with predictors for within-group pairs and group-specific effects (movie–group combinations) while controlling for an intercept capturing changes common to all participants (e.g., re-watching effects). Across-groups analyses used a single predictor marking same-group pairs; group-specific analyses used group indicator matrices. For novel clips, post-conversation ISC was modeled using the same predictors to test generalization. Significance used subject-wise permutation testing (rows/columns permuted identically) with cluster-level correction (AFNI 3dClustSim; non-Gaussian ACF). Rolling-window ISC (10 TR window, step 1 TR) characterized time courses in significant clusters. Neural influence analysis: For each ordered ego–alter pair, neural influence quantified how much the alter’s post-conversation BOLD time series moved toward the ego’s pre-conversation time series relative to pre-conversation similarity. Whole-brain neural influence was the sum of positive values in unthresholded influence maps. Multiple regressions related neural influence to social network centrality (eigenvector, brokerage, PCA centrality) of egos and alters, with mixed-effects models used for whole-brain influence. Conversation coding: Transcripts were coded by hypothesis-blind raters for speech-turn types and continuous properties. Perceived social status was rated by additional coders; inter-rater reliability was high for status and moderate for speech-turn coding. At the group level (n = 45 group–movie pairs), hierarchical models (lme4) related neural and behavioral alignment to words spoken, inequality (Gini) of PCA centrality and perceived status, and their interactions, with random intercepts for movie. Social network measures: Cohort-wide name-generator survey created directed social graphs; eigenvector centrality and brokerage (inverse of constraint^0.5) were computed (igraph). PCA centrality captured variance shared by eigenvector and brokerage. Participants in the fMRI study spanned a broad range of centrality values.

• Post-conversation behavioral convergence: Participants reported agreement with group consensus (scale −3 to +3) with mean = 1.71; t(28) = 8.32, p < 0.001, 95% CI [1.29, 2.13]. Survey answers became more similar within groups after conversation; hierarchical model explained behavioral distance (marginal R² = 0.25, permutation p < 0.001, n = 136 participant pairs), with significant effects of session, comparison type, and their interaction.
• Increased neural alignment after conversation: ISC increased within conversation groups across visual and auditory cortices and higher-order regions (TPJ, angular gyrus, posterior cingulate, medial prefrontal cortex, temporal pole). Control participants (no conversation) showed decreases or no localized positive changes. Group- and movie-specific analyses (movie–group combinations; 45 possible) indicated that the loci of alignment depended on who spoke and what was discussed.
• Generalization to novel stimuli: For previously unseen clips from the same movies, within-group alignment was higher in conversation groups, including bilateral anterior frontal gyrus across groups, and broader networks at the movie–group level, indicating that conversation provided a framework for interpreting new information.
• Behavioral–neural link: Similarity of survey answers correlated with whole-brain alignment across all participant pairs (including controls): r(3476) = 0.09, permutation p < 0.001; mixed-effects standardized β = 0.11, 95% CI [0.07, 0.13], p < 0.001.
• Neural influence and social network centrality: Mixed-effects model predicting whole-brain neural influence from eigenvector and brokerage centrality (egos and alters) had marginal R² = 0.03, permutation p = 0.001; eigenvector centrality (alter) was a significant positive predictor (β = 0.43, 95% CI [0.12, 0.74], p = 0.007). PCA and eigenvector centrality of alters tended to predict positive neural influence in widespread areas; ego centrality sometimes predicted negative influence in specific regions, suggesting distinct processes captured by each measure.
• Conversation dynamics, status, and alignment: Groups that spoke more had higher neural alignment (β ≈ 0.56, p < 0.001); groups with higher-centrality participants had higher neural alignment (β ≈ 0.52, p < 0.001). Words spoken moderated the centrality effect (interaction β ≈ 0.32, p = 0.04). Unequal turn-taking predicted lower neural alignment (β ≈ −0.35, p ≈ 0.037).
• Behavioral alignment at group level: Groups with higher centrality participants showed higher behavioral alignment (β = 7.24, 95% CI [4.15, 10.33], p < 0.001). However, when high-centrality co-occurred with unequal perceived social status, behavioral alignment was much lower (β = −17.27, 95% CI [−24.77, −9.68], p < 0.001).
• Perceived status versus centrality: Higher perceived status correlated with more words spoken and higher influence ratings by peers, yet was negatively correlated with pairwise whole-brain neural influence (r(208) = −0.05, p < 0.001; mixed-effects standardized β = −0.05, 95% CI [−0.085, −0.016], p = 0.005). Centrality was not significantly correlated with amount spoken (r(37) = 0.24, p = 0.143) or influence ratings (r(27) = 0.15, p = 0.433). Qualitative transcript analyses indicated high-centrality participants encouraged others to contribute, whereas perceived high-status participants more often challenged or implicitly rejected others’ proposals, corresponding to greater public compliance but lower private acceptance.

The findings support the hypothesis that consensus-building conversation can align neural processing among group members beyond mere public compliance. After discussion, groups exhibited increased ISC in sensory and higher-order brain regions associated with attention, default mode, language, memory, and social cognition, and this alignment generalized to novel, previously unseen clips from the same movies. The correlation between behavioral convergence and whole-brain neural alignment suggests that converging interpretations underlie shared neural dynamics. Variation in conversational success was informative: Groups with participants perceived as high social status tended to exhibit unequal turn-taking, more orders and rejections of others’ proposals, and lower neural alignment, consistent with public compliance without private acceptance. In contrast, real-world social network centrality was associated with reciprocal neural influence—central individuals became more similar to others and others became more similar to them—and with higher group neural and behavioral alignment. These patterns imply that socially central individuals may facilitate consensus and cognitive alignment by eliciting broader participation and integrating diverse viewpoints. Overall, the results connect conversational behavior and social network structure to shared neural representations of narrative content.

This study demonstrates that naturalistic, consensus-seeking conversation increases within-group neural alignment and that such alignment can generalize to novel but related stimuli. Behavioral convergence correlates with neural convergence, indicating conversation shapes not only expressed agreement but also shared cognitive processing. The work differentiates the roles of perceived social status versus real-world network centrality: perceived status is linked to conversational dominance and lower private neural alignment, whereas network centrality relates to mutual neural influence and higher group alignment. These insights have implications for understanding social influence, group decision-making, and the diffusion of information through networks. Future research should identify mechanistic pathways by which conversation produces neural alignment, test generalization across broader stimuli and populations, and examine how specific conversational behaviors can be leveraged to promote constructive consensus in real-world settings.

The reported effects are relatively small and the statistical generalization is limited: movies and groups were treated as fixed effects, so results may not generalize to other clips or group compositions. Small group sizes and conservative multiple-comparison corrections may have reduced sensitivity to some effects. Control analyses indicate that simply re-watching is insufficient to explain the observed changes, but unmeasured factors could contribute. More broadly, while conversation-induced neural alignment was observed, the mechanistic details remain unknown. The sample is a specific population (MBA students at a rural private university), which may limit external validity. Some reported confidence intervals and model details exhibit irregularities, and exploratory conversation-level models may be sensitive to model specification and measurement noise.