Basolateral and central amygdala orchestrate how we learn whom to trust

The study addresses how humans learn to distinguish trustworthy from untrustworthy interaction partners and how this learning is implemented in amygdala subnuclei and connected subcortical regions. Building on rodent work and prior human lesion evidence indicating a critical role of the basolateral amygdala (BLA) in social experiential learning, the authors note that conventional human neuroimaging often treats the amygdala as a unitary structure, potentially obscuring functional specialization of subnuclei (BLA vs central amygdala, CeA). Using an fMRI protocol optimized for ventral brain regions and time-resolved analyses, the study aims to clarify phase-specific roles of BLA and CeA across distinct trust game phases (preparation, investment, waiting, outcome), how these signals evolve with learning over time, and how they relate to behavior and subjective trustworthiness beliefs. Secondary, more exploratory aims examine the involvement of nucleus accumbens (NAc), substantia nigra/ventral tegmental area (SN/VTA), bed nucleus of the stria terminalis (BST), and basal forebrain, given their connectivity with the amygdala and roles in reward, motivation, and uncertainty processing.

Prior work in rodents implicates the BLA in learning from social experiences and differentiating positive and negative associations. Human lesion studies show that selective bilateral BLA damage impairs adaptation of trust behavior to trustworthy vs untrustworthy partners in repeated trust games. However, neuroimaging in neurotypical humans has yielded inconsistent amygdala findings, likely due to limited ability to resolve subnuclei. The amygdala’s structural and functional heterogeneity suggests distinct roles for BLA (sensory integration and outcome evaluation) and CeA (motivational/affective processes), with differing connectivity profiles: BLA to sensory integration and lateral PFC, CeA to ventral striatum/NAc and medial PFC. The extended amygdala, including BST, is linked to sustained stress, unpredictability, and social cognition, while NAc and dopaminergic midbrain (SN/VTA) support reward learning and social decision-making. Rodent studies indicate BLA–NAc pathways mediate reward learning and NAc excitability depends on BLA input, suggesting coordinated roles in reward-seeking and trust evaluation.

Participants: The main sample comprised 62 healthy, neurotypical volunteers (mean age ~23.8 ± 3.15 years). All provided informed consent; procedures were approved by the Medical University of Vienna ethics committee (EK-Nr. 148/2015). Participants were mostly undergraduate students and screened for standard MRI safety and exclusion criteria. Task: Participants performed a repeated trust game in the MRI scanner with two ostensible partners: one trustworthy and one untrustworthy. In reality, both partners were simulated. Across 40 rounds (2 × 20 per partner type), participants always acted as investor. Each round involved four phases: (1) Preparation (face of current partner), (2) Investment (choose amount from endowment of 10 units to invest; investment tripled for the partner), (3) Waiting (partner’s decision), (4) Outcome (partner’s back-transfer). In the first two rounds, both partners returned the invested amount. Subsequently, the trustworthy partner returned as much or more than invested; the untrustworthy partner returned less than or equal to the investment. Investment amounts indexed behavioral trust. After the task, participants rated the partners on trustworthiness, attractiveness, likeability, and reliability (VAS 1–10). MRI acquisition: 3T Siemens scanner; optimized MB-EPI for ventral brain, TR = 2,160 ms, TE = 32.2 × 2,33 ms, 96 × 92 × 20 voxels, 1.5 mm isotropic. Preprocessing and analysis: SPM12 used for realignment, coregistration, normalization to MNI space (1.5 mm isotropic), and smoothing (6 mm FWHM). First-level GLMs modeled the four task phases for each partner type, with motion regressors and high-pass filtering (cutoff ~1/128 Hz). Region-of-interest and time-resolved analyses focused on anatomically defined BLA and CeA volumes of interest; functional connectivity profiles were examined to validate subnuclei differentiation. Secondary ROI analyses evaluated NAc, SN/VTA, BST, and basal forebrain. Group-level analyses tested phase- and partner-dependent effects, and relationships with behavior and post-task ratings. Learners vs non-learners were defined via a median split on the investment difference between trustworthy and untrustworthy partners (Δ investment).

Behavioral: - Participants invested more in the trustworthy than the untrustworthy partner overall. Investment differences (Δ investment) correlated with post-experiment trustworthiness ratings (r ≈ +0.39, p = 0.002). - Learners (median split on Δ investment) adapted to favor the trustworthy partner over time; non-learners showed reduced differentiation. Phase-specific amygdala subnuclei responses (Bonferroni-corrected): - BLA: No significant activation in preparation; significant deactivation during waiting and significant activation during outcome evaluation (e.g., outcome: PSC ± 95% CI ≈ -0.11 ± 0.02, t ≈ 9.8, p < 0.0001). - CeA: Significant activation during preparation (PSC ± 95% CI ≈ 0.04 ± 0.02, t ≈ 3.2, p = 0.0021); significant deactivation during investment and waiting; no significant effect during outcome. - Across the full sample, no overall BLA/CeA differences between trustworthy vs untrustworthy partners in any phase. Learners vs non-learners: - Only in learners did BLA and CeA show greater activation for the untrustworthy vs trustworthy partner during the preparation phase, and these differences correlated with both investment differences and subjective trustworthiness ratings. Secondary subcortical regions during outcome (trustworthy > untrustworthy): - NAc: t = 6.73, p < 0.0001 - SN/VTA: t = 3.37, p = 0.0005 - BST: t = 4.00, p = 0.0002 - Basal forebrain: t = 2.1, p = 0.0361 - These effects were present irrespective of learning group, indicating outcome-related reward/motivation processing did not differentiate learners from non-learners. Connectivity: - BLA connectivity predominated with sensory integration and lateral PFC; CeA with ventral striatum (including NAc) and medial PFC, supporting functional dissociation.

Findings reveal distinct, time-resolved roles of amygdala subnuclei in trust learning and decision-making. The BLA was maximally engaged during outcome evaluation, consistent with encoding and updating beliefs about others’ trustworthiness based on reciprocation outcomes. Crucially, BLA and CeA signals during the preparation phase differentiated untrustworthy from trustworthy partners only in learners and predicted both subsequent trust behavior (investments) and subjective trustworthiness beliefs, indicating their role in guiding behavior via anticipatory evaluation. In contrast, reward-related processing during outcome, reflected in greater responses in NAc, SN/VTA, and BST to trustworthy outcomes, did not differ between learners and non-learners, suggesting that such outcome valuation mechanisms are not the primary driver of individual differences in trust learning. CeA activity appears linked to motivational and affective valuation during preparation, covarying with subjective beliefs rather than directly controlling investment behavior. Together, the results bridge animal and human evidence, demonstrating that amygdala subnuclei orchestrate learning about whom to trust, with BLA supporting belief updating and discrimination of partner trustworthiness, CeA supporting motivational-affective preparation, and connected regions (NAc, BST, SN/VTA, basal forebrain) contributing to reward and uncertainty processing during outcomes.

This study demonstrates that human amygdala subnuclei play dissociable roles in learning whom to trust: BLA supports outcome-based belief updating and discrimination between trustworthy and untrustworthy partners, while CeA supports motivational-affective preparation linked to subjective trustworthiness beliefs. Reward-related outcome processing in NAc, SN/VTA, and BST was stronger for trustworthy outcomes but did not distinguish learners from non-learners, underscoring that trust learning hinges more on amygdala engagement than on general reward encoding. The work translates lesion and animal-model insights to neurotypical humans using time-resolved, high-resolution fMRI focused on amygdala subnuclei. Future research should combine variants of the trust game with computational modeling to dissect learning signals (e.g., prediction errors, volatility) across amygdala–striatal–midbrain circuits and test causal roles (e.g., via neuromodulation or patient studies).