Callousness, exploitativeness, and tracking of cooperation incentives in the human default network

Significance

To navigate complex social lives, humans must learn who is worth cooperating with and whom to avoid. Here, we show that humans learn to predict the behavior of social counterparts by integrating reputational information with first-hand experience of social interactions. We find that superior learning during social interactions is related to compassion (vs. callousness) and associated with the activity of the brain’s default network. Remarkably, however, cooperation-related learning signals in the default network are also associated with self-reported manipulativeness and exploitativeness, suggesting that the ability to anticipate others’ choices may serve both the light and dark sides of human behavior.

Abstract

Human cognitive capacities that enable flexible cooperation may have evolved in parallel with the expansion of frontoparietal cortical networks, particularly the default network. Conversely, human antisocial behavior and trait antagonism are broadly associated with reduced activity, impaired connectivity, and altered structure of the default network. Yet, behaviors like interpersonal manipulation and exploitation may require intact or even superior social cognition. Using a reinforcement learning model of decision-making on a modified trust game, we examined how individuals adjusted their cooperation rate based on a counterpart’s cooperation and social reputation. We observed that learning signals in the default network updated the predicted utility of cooperation or defection and scaled with reciprocal cooperation. These signals were weaker in callous (vs. compassionate) individuals but stronger in those who were more exploitative (vs. honest and humble). Further, they accounted for associations between exploitativeness, callousness, and reciprocal cooperation. Separately, behavioral sensitivity to prior reputation was reduced in callous but not exploitative individuals and selectively scaled with responses of the medial temporal subsystem of the default network. Overall, callousness was characterized by blunted behavioral and default network sensitivity to cooperation incentives. Exploitativeness predicted heightened sensitivity to others’ cooperation but not social reputation. We speculate that both compassion and exploitativeness may reflect cognitive adaptations to social living, enabled by expansion of the default network in anthropogenesis.

More than any other species, humans’ social lives depend on flexible cooperation. Anthropoids began living in loose social groups to reduce predation risk over 50 Mya, and hominins have evolved increasingly sophisticated forms of sociality more recently (1). This transition to sociality conferred major advantages but also exposed individuals to intergroup rivalries (2). Selection pressures thus favored social cohesiveness, with individuals forging cooperative bonds to reduce conflict and coordinate foraging, rearing, and defense (3). The computational demands imposed by this social environment may have contributed to the recent expansion of the human cortex, particularly the default network (4). It is no secret, however, that in cooperative societies, many people act selfishly, even aggressively. In fact, frequency-dependent selection often sustains noncooperative minorities among largely cooperative populations (5), with noncooperation remaining advantageous as long as it is rare (6). Indeed, variability in human antisocial behavior is heritable (7); it is captured by a trait known as antagonism v. agreeableness in dimensional models of personality (8).

Agreeableness is the tendency to coordinate one’s goals with those of others to cooperate effectively and to avoid hostility (9, 10). Its opposite pole, antagonism, is a central feature of personality disorders that predicts difficulties with attachment and intimacy; aggressive, hostile, and criminal behavior toward others; and, in treatment, fractures in the therapeutic relationship (11–13). Agreeableness–antagonism consists of two lower-order traits. Compassion–callousness contrasts empathy and generosity with indifference and cruelty; callousness is a key, heritable component of childhood conduct disorder and adult antisocial personality (14, 15). Politeness–exploitativeness contrasts honesty and humility with selfishness and manipulativeness (16), central features of narcissism. Interestingly, whereas callousness is associated with inferior social cognition, the opposite is true of exploitativeness (17), raising the question of whether the two traits are related to dissociable and even opposing neural processes that control cooperation during social interactions.

Humans cooperate more willingly when they anticipate reciprocation and when cooperation provides benefits relative to noncooperation (c.f. responsiveness in game theory and reciprocal altruism in ethology; 18–20). Conversely, cooperation with defectors is costly, and even more so when they coax the individual to cooperate but fail to reciprocate (21, 22). Hence, in most settings, the Nash equilibrium strategy is to always defect. Consequently, the neural computations supporting cooperation likely encode the expected advantage of choosing to cooperate or defect, based on a prediction about whether the other party will also cooperate (23). Such predictions can be influenced by a counterpart’s prior reputation conveyed by third parties, as demonstrated in human neuroimaging studies showing that reputation biases striatal activity during economic exchanges (24, 25). More recently, we found that learning signals in the striatum and default network track the expected return (value) of one’s interpersonal policy.^* These regions’ activity increased following both correctly anticipated cooperation and correctly anticipated defection (27). This pattern matches the intuition that it is rewarding to predict not only a counterpart’s cooperative actions but also their defections, as when one elects not to ask out an acquaintance and later learns from a mutual friend that the acquaintance was never interested. Our findings agree with studies describing how individuals incorporate a social counterpart’s intentions when deciding to reciprocate (28–31), a process that relies on the default network (32, 33).

The default network spans the medial prefrontal cortex, posterior cingulate cortex (PCC), inferior parietal lobule, and medial and lateral temporal regions (34, 35). It expanded disproportionately in anthropogenesis (36, 37), in part due to the expression of genes that diverge with the human lineage and are associated with sociability (38). In humans, the structure and connectivity of the default network scale with social network size (39). Importantly, default network activity has been linked to high agreeableness (low antagonism) and better performance on social cognition tasks (17, 40). These observations suggest that phylogenetically recent circuits within the human default network reflect adaptations to sociality and are a likely substrate of individual differences in cooperativeness and broader social behavior.

Building on these ideas, we investigated the neural signatures of the facets of agreeableness–antagonism, compassion–callousness and politeness–exploitativeness, in a sample of participants with (n = 113) and without (n = 55) a probable or definite diagnosis of borderline personality disorder (BPD), which ensured that we captured the full range of severity in the latent trait constructs of interest. Whereas prior studies of the default network have examined its structure, resting-state connectivity, and activation to social stimuli, we focused on its role in learning to predict the return on one’s cooperation. Specifically, we used a computational model of learning during social exchanges, functional neuroimaging, and multilevel structural equation modeling (MSEM) to examine how cooperation-related learning signals in the default network are related to behavioral reciprocation, callousness, and exploitativeness. Our earlier behavioral findings (17) led to the prediction that while callous individuals would show insensitivity to a counterpart’s reputation or actual reciprocation, exploitative individuals would be more sensitive to a counterpart’s cooperation. Further, we expected that learning signals tracking the value of one’s current interpersonal policy, encoded in the default network, would be negatively related to callousness and positively related to exploitativeness, and would account for the two facets’ distinct effects on behavioral reciprocity.

Results

Modified Trust Game: Behavior.

In social exchanges, participants integrate reinforcement with reputational information in an additive manner to make a prediction about whether the trustee will return an investment, which governs their decisions to cooperate (27). Based on prior work, we expected that callousness would be related to a broad insensitivity to both influences, whereas exploitativeness would be associated specifically with a greater sensitivity to reinforcement history. To evaluate these predictions, it was critical that we independently manipulated the counterpart’s cooperation rate and social reputation. To this end, participants interacted with three different trustees of varying reputations (good, bad, neutral; 48 trials each, 144 trials total). On each trial, participants received $1.00 to keep or invest (i.e., share) with the trustee. If they chose to keep, they retained their dollar. If they chose to invest, the outcome was contingent on the trustee: If the trustee returned, they received $1.50, whereas if the trustee kept, they were left with nothing (Fig. 1B). All trustees returned 50% of the time in the first block of 16 trials (allowing us to observe the effect of reputation independent of reinforcement), and then either 25% (poor block) or 88% (rich block) of the time in two subsequent counterbalanced blocks of 16 trials. Critically for our analysis of individual differences, participants were shown the trustee’s decision even after choosing to keep the $1.00. This ensured that all participants experienced the same feedback, including when they predicted a defection and regardless of their own cooperation rate.

Fig. 1.

Rationale, design, and analytic approach. Individuals learn from experience by selecting an action, observing its outcome, and updating the expected reward value of future actions. Value updates are made using PEs, which reflect the discrepancy between expected and obtained outcomes such that better-than-expected outcomes lead to positive PEs and worse-than-expected outcomes lead to negative PEs. Other salient social information can also be integrated with experience to influence social decision-making, as when the reputation of one’s partner predicts decisions to trust them even when reputation is unrelated to the partner’s actual behavior (27, 41). (A) Consider the decision about whether to buy a friend a holiday gift. Reputational information (e.g., news of a friend’s immoral behavior; blue circle) can be integrated with reinforcement history (e.g., did the friend buy you a gift last year? green circle) to affect one’s policy toward their social counterpart. Critically, decisions that correctly anticipate the behavior of one’s social partner (correctly predicting they bought you a gift, or correctly predicting they did not buy you a gift) yield positive PEs according to our policy model, leading to a cycle of reciprocity even when no actual reward is received. (B) Participants played a modified iterative social trust game with three fictional trustees, in which they had the option to keep an initial endowment or invest it in the hopes of increasing their profit if the trustee also invested. Pretask vignettes were used to manipulate trustees’ reputations (blue box). To manipulate reinforcement history, trustees returned at varying rates across rich, poor, and neutral blocks (green box). On trials where participants kept, counterfactual feedback about what the trustee would have chosen was provided, even though it did not affect the trial payout. (C) The policy model posits that individuals learn from social feedback to optimize their approach, or policy, toward their social counterpart, leading to better anticipation of the counterpart’s behavior. One implication of this is that correct predictions of the trustee’s behavior will lead to positive reward PEs, even if no actual reward is provided. This can be seen by comparing the expected direction of PEs in a model in which participants track actual rewards (column 3) vs. a model in which they track the success of their policy toward the counterpart (column 4). We propose that policy PEs are primarily encoded within the brain’s default network. Image credit: Default network figure adapted from ref. 35. (D) MSEM was used to evaluate whether between-person variables (e.g., policy PEs encoded in the default network) moderate the effect of design variables on trial-level decision-making. Personality traits were introduced as between-person predictors of policy PEs in the default network. Formal tests of mediation were then used to examine the indirect effect of traits on behavior via learning signals (highlighted lines).

Participants were less likely to invest with the bad trustee than neutral or good trustees (bad v. neutral: β = −0.24, SE = 0.05, P < 0.001; bad vs. good: β = −0.26, SE = 0.05, P < 0.001). There was also a reciprocity effect (β = 0.62, SE = 0.07, P < 0.001) such that when trustees returned on one trial, participants often invested on the next and vice versa (full behavioral results in SI Appendix, section IV). Importantly, participants varied greatly in how they decided whether to cooperate (SI Appendix, Fig. S2), enabling our subsequent analyses of individual differences. Because the rate of cooperation did not vary between the good and neutral trustees, these conditions were combined for subsequent analyses.

Computational Model of Learning during Social Exchanges.

We used reinforcement learning to model how participants learn from the trustee’s previous decisions to cooperate or defect (Fig. 1C). Our model also captured individual differences in participants’ rate of unconditional cooperation and their sensitivity to the reputation of the trustee. We have previously found that participants learn from actual rewards as well as counterfactual outcomes, using their own alternative action as the reference, an algorithm we term “the policy model” (27). To replicate that observation and select the best model for our neural analyses, we compared the policy model to three alternative models, which learned from i) only actual rewards, ii) regret signals, with the best possible outcome being the reference, or iii) actual rewards and counterfactual outcomes with trustee’s alternative action as the reference. Thus, each model had a distinct payoff matrix (SI Appendix, section VI). All four models included the same five subject-specific parameters: $\alpha$, a learning rate parameter reflecting the speed at which participants update the value of keeping or investing; β, a temperature parameter reflecting choice stochasticity, κ_subject, a participant-level parameter reflecting a bias of the participant to keep or invest regardless of reinforcement; and two condition-level κ_trustee parameters, reflecting the participant’s bias to keep or invest with the good or bad trustee relative to the neutral trustee (note that κ_trustee(good) is zero during exchanges with the bad trustee, and vice versa). In addition, to examine whether reputation moderated learning as opposed to directly affecting investing, we tested a variant of the policy model that included trustee-specific learning rates and a single participant-level bias parameter (to reflect a general bias to invest). Bayesian model comparison (BMC) of the five competing models indicated that the policy model with a single learning rate dominated all alternatives (Bayesian omnibus risk < 0.001; protected exceedance probability for the policy model = 1.00; see model frequencies in SI Appendix, Fig. S5D), replicating our previous findings (27). Parameters from the winning policy model were minimally intercorrelated (SI Appendix, Fig. S5A).

Next, we examined whether participant-level parameters derived from the policy model related to individual differences in antagonism and its two facets, callousness and exploitativeness (see SI Appendix, section VI for full results). Traits were derived using exploratory factor analysis (EFA) (see Fig. 2 and SI Appendix, section V for details). Antagonism was negatively related to the learning rate (β = −0.05, SE = 0.02, P = 0.02), suggesting that antagonistic individuals were slower to update keep/invest values based on additional information (SI Appendix, Fig. S5B). Callousness (controlling for exploitativeness) was positively associated with the κ_trustee parameter for the bad trustee (β = 0.05, SE = 0.02, P = 0.03; SI Appendix, Fig. S5C). In general, participants tended to invest less with the bad trustee than the neutral trustee, but this effect was abolished in those high on callousness. This pattern is illustrated by our model-free behavioral results in which callousness, but not exploitativeness, moderated the effect of reputation on investing (β = 0.26, 95% CI = [0.04, 0.45], P = 0.01; SI Appendix, Table S14). Participants who were low (−1 SD) or average on callousness were less likely to invest with the bad trustee relative to the good and neutral trustees, whereas reputation had no effect on investing for those high (+1 SD) in callousness (SI Appendix, Fig. S8C).

Fig. 2.

Two-factor solution to an exploratory factor analysis of antagonism-agreeableness scales.

Policy PE Signals in the Default Network.

A previously proposed model of cooperation tracked trustee’s social value, predicting positive updates when trustees cooperate and negative updates when they defect (42). However, we have shown that striatal and default network signals are better described by a model that tracks the value of one’s current interpersonal policy (the distribution of cooperation/defection probabilities conditional on a trustee’s reputation and actual cooperation), predicting positive PEs following correctly anticipated defections by the trustee [(27), SI Appendix, Fig. S7]. That is, both unexpected cooperation and defection generate policy PEs, matching the agent’s cooperation rate to the counterpart’s predicted cooperation rate. Based on our previous results and other studies implicating the default network in predicting others’ behavior, we expected to find policy PE signals both in the canonical striato-limbic reward PE network and in the default network. Fig. 3A depicts the whole-brain map (FWE P < 0.05) with signed policy PEs found in the canonical network of the ventral and dorsal striatum, and putamen, as well as in the lateral occipital cortex, pre- and postcentral gyrus, cerebellum, and the frontal operculum. Policy PEs were also prominent throughout the default network, and especially in the posterior hub, with whole-brain significant PE activation in 85% of voxels contained in an established definition of this region (43). Policy PEs were also present in other regions of the default network, including the anterior cingulate cortex and the bilateral middle and inferior temporal gyrus (SI Appendix, Table S6 provides a complete listing of clusters). Overlap between policy PEs and the canonical default network is depicted in Fig. 3B.

Fig. 3.

(A) Whole-brain policy PE map. Threshold z > 5.17, p_ptfce < 0.05. (B) Overlap between surviving PE clusters and regions of the canonical default network, broken down by subsystem. Note that 85% of the voxels in the posterior hub of the default network were statistically significant. Images created using ggbrain (44).

Policy PE Signals in the Default Network and Reciprocity.

Since learning from counterfactual feedback facilitates reciprocal cooperation, we expected policy PEs to be associated with greater reciprocity. We tested this prediction in a multilevel structural equation model (MSEM) where individual differences in the strength of default network PE responses predicted the effects of exchange (trial within trustee), reputation, and trustee’s previous decision on participant decisions to track the value of one’s current interpersonal policy (SI Appendix, Table S7). Individuals with stronger PE signals in the default network exhibited greater reciprocity (β = 0.25, 95% CI = [0.09, 0.39], P < 0.001). Specifically, for those with high, but not low, PE signaling in the default network, the trustee’s decision to return on the previous trial was associated with a higher likelihood of the participant reciprocating (Fig. 4A). PE signaling in the overall default network had no effect on the association between reputation and the participant’s decision to invest.

Fig. 4.

(A) Policy PEs in the default network moderate the effect of the trustee’s last decision on the participant’s choice to keep or invest. (B) Policy PEs in the medial temporal subsystem moderate the effect of trustee reputation on participant decisions. (C) Callousness was negatively, and (D) exploitativeness was positively, associated with policy PEs in the overall default network. Abbreviations: PE = prediction error; DN = default network; MT = medial temporal subsystem; SD = standard deviation.

The default network is functionally heterogeneous, spanning three subsystems: core, medial temporal, and dorsal medial (35). As a test of anatomical specificity, we examined whether the effect of policy PE signals on behavior was driven by a particular subsystem. A follow-up MSEM indicated that the effect of reciprocity on sharing was not moderated by PE signals within any single subsystem (SI Appendix, Table S8), suggesting that the effect of learning on behavior occurs at the network level. Notably, however, the effect of reputation on participants’ decisions to keep or invest was moderated by the medial temporal subsystem, even after controlling for the core and dorsal medial subsystems (β = −0.31, 95% CI = [−0.53, −0.03], P = 0.04). Specifically, the likelihood of sharing was lower with a bad (vs. good, neutral) trustee only for those with stronger PE signals in the medial temporal subsystem (Fig. 4B).

Antagonism, Callousness, Exploitativeness, and Policy PEs in the Default Network.

If policy PEs enable individuals to learn about and correctly anticipate the behavioral strategies of others, they should be negatively related to callousness and positively related to exploitativeness. Indeed, controlling for age and sex, we found that callousness was negatively (β = −0.39, SE = 0.10, P < 0.001) associated with PE signaling in the default network, whereas the reverse was true for exploitativeness (β = 0.26, SE = 0.10, P = 0.01; Fig. 4 C and D and SI Appendix, Table S9). Conversely, higher-order antagonism was unrelated to PEs (β = −0.10, SE = 0.08, P = 0.22). These effects held across all three subsystems of the default network (SI Appendix, Fig. S7; the effect of exploitativeness on PE signaling in the dorsal medial subsystem is marginal but qualitatively similar, P = 0.05). Exploratory analyses testing the anatomical specificity of these relationships indicated that policy PE signaling in the dorsal attention network and striatum were unrelated to exploitativeness, but negatively related to callousness (dorsal attention: β = −0.24, 95% CI = −0.41, −0.06, P = 0.01; striatum: β = −0.25, 95% CI = −0.41, −0.07, P = 0.01; see SI Appendix, section XV). This pattern of results suggests that antagonism’s two facets differentially predict PE signaling in the default network. For callousness, this effect also extended to other regions and networks involved in social learning, suggesting a brain-wide dampening of social PE signals.

Effects of Callousness and Exploitativeness on Behavior Depend on Policy PE Signaling in the Default Network.

Were differential effects of callousness vs. exploitativeness on cooperation-related learning signals related to behavioral reciprocity? To answer this question, we extended the MSEM reported above by allowing the two facets of antagonism to predict default network policy PE signaling and tested whether PE signaling mediated the effect of callousness and/or exploitativeness on reciprocity. Both callousness and exploitativeness had a significant indirect effect on reciprocity via PE signaling (see Fig. 5 and SI Appendix, Table S10 for full model results; callousness: b = −0.05, 95% CI = −0.10, −0.02, P < 0.001; exploitativeness: b = 0.03, 95% CI = 0.004, 0.07, P = 0.03). These results indicate that blunted policy PEs within the default network account for the association between callousness and lower reciprocal cooperation. In contrast, amplified policy PE signals serve as the link between exploitativeness and greater reciprocity. Notably, these effects are limited to callousness and exploitativeness, as PE signaling in the default network did not mediate the effect of higher-order antagonism on behavior (SI Appendix, section XIV). Moreover, the effect is specific to PE signals in the default network: Policy PEs in the striatum and dorsal attention network did not account for the association between traits and behavior (SI Appendix, section XV). Finally, in sensitivity analyses, the effects of callousness and exploitativeness on reciprocity via policy PEs were evident even after controlling for impulsivity, intellectual functioning, and negative affect (SI Appendix, Tables S18–S20). Critically, participants with a probable or definite diagnosis of BPD were higher on both callousness and exploitativeness than healthy comparison participants (Mann-Whitney U test for callousness: W = 710, P < 0.001; exploitativeness: W = 1253, P < 0.001; see SI Appendix, Table S22 for means and SDs). When BPD status was added as a covariate to our models, the effects of callousness and exploitativeness on reciprocity via policy PEs held, suggesting that trait variation, as opposed to diagnostic status, is the primary driver of our observed effects (SI Appendix, Table S21).

Fig. 5.

Multilevel structural equation model depicting the effects of callousness and exploitativeness on reciprocity via PE modulation of the default network. Boxes reflect observed variables. Circles reflect latent variables. Filled black circles indicate random slopes. Dotted lines depict indirect effects. Only significant paths are shown. Exchange number (i.e., trial within trustee, a within-person predictor), age, and sex covariates are excluded for parsimony. All coefficients are standardized except the indirect effects. *P < 0.05.

Discussion

The evolution of human cooperation has been linked with the expansion of frontoparietal networks, and particularly the default network (38). We argue that cognitive capacities enabled by this expansion and traditionally thought to subserve prosociality are also related to the personality traits of manipulativeness and exploitativeness, representing the dark side of human nature. Using reinforcement learning modeling and fMRI, we investigated how humans learn to predict a counterpart’s choices during a modified iterative trust game. We observed that reciprocal cooperation in this context depends on learning signals in the default network tracking updates to the expected return on one’s own cooperation, based on the counterpart’s cooperation history (i.e., policy PEs). More robust cooperation-related learning signals in the default network were seen in people who were compassionate and empathic, but strikingly, also in people who were manipulative and exploitative. These learning signals, which represent experience-driven changes in the counterpart’s estimated cooperation rate, accounted for the association between each trait and reciprocity in a social exchange, consistent with the idea that the human default network facilitates the flexible navigation of social interactions.

In this study, both compassion and exploitativeness were associated with an enhanced capacity to learn from social experience, yielding increased reciprocity and social coordination in a trust game. It is possible that for those high in compassion, superior learning facilitates social connections and interdependence. Although lower reciprocation in callous people may seem disadvantageous, by defecting after the counterpart cooperates, they prevent deceitful counterparts from coaxing them into cooperation, which is rational in a dog-eat-dog world where others deceive and defect rather than reciprocate. In contrast, more exploitative people may use learning strategically, eliciting cooperation that both serves their immediate self-interest and creates opportunities for long-term exploitation. Indeed, the ability to strategically manipulate, persuade, or exploit others has been associated with greater achievement and satisfaction in certain contexts (45). Similarly, components of exploitativeness such as attention-seeking, immodesty, or exhibitionism overlap with agentic extraversion (46), which some studies link with enhanced reward learning (47). These traits also characterize grandiose narcissism, suggesting that default network learning mechanisms that enable one to track incentives for cooperation are also important for asserting dominance and pursuing status. Nonetheless, while our results suggest that learning to anticipate another’s behavior is associated with self-reported exploitativeness, our experiment did not capture interpersonal exploitation directly. Future studies are needed to investigate how learning-based predictions are utilized to manipulate and exploit others.

Whereas callousness and exploitativeness were both related to reciprocity on the game via default network sensitivity to policy PEs, only callousness dampened the effect of reputation. Callousness may reflect an insensitivity to social signals and cues, as evidenced by studies linking it with blunted electrophysiological responses to emotional, particularly fearful, faces (48, 49). Highly conserved neural mechanisms of empathy are thought to have emerged as adaptations to infant care, pair bonding, and shared vigilance (10, 50). In rodents and primates these behaviors depend on subcortical circuitry sensitive to sex hormones and neuropeptides, such as the anterior hypothalamus and medial amygdala (part of the social behavior network; 51). This circuitry predates the expansion of frontoparietal networks and lateral temporal regions that enabled more advanced primate social cognition (the cognitive social brain, see review by ref. 52).

Our findings link both callousness and exploitativeness with cooperation-related learning signals in the default network during social exchanges. Consequently, it is interesting to consider whether the distinct behavioral profile of these two facets could reflect learning-related inputs from the subcortical social behavior network to the default network. For example, default network lesions in early Alzheimer’s disease do not usually result in socially inappropriate or callous behavior, unlike changes in amygdala or hippocampus, which are observed both in antagonism and in neurodegenerative disease with social syndromes (53–55). Consistent with this conjecture, learning signals in the striatum and the evolutionarily conserved medial temporal subsystem of the default network (37) scaled with callousness, but not exploitativeness. Learning signals in the medial temporal subsystem also scaled with behavioral sensitivity to reputation based on counterparts’ biographical depictions. These findings are consistent with the role of parahippocampal and retrosplenial cortices in retrieving contextual information and episodic detail (56), and with evidence of hippocampal alterations in callousness and antisocial behavior (53, 55); thus, callous individuals may fail to integrate episodic memories and contextual information about others when deciding whether to cooperate.

The posterior hub of the default network may be critical to integrating contextual information encoded in subcortical circuitry with information learned from social exchanges to inform cooperation decisions. Among the canonical default network regions, PCC, retrosplenial cortex, and precuneus showed the most widespread sensitivity to unpredicted cooperation and defection outcomes (85% of the voxels in the posterior hub of the default network were statistically significant). Dorsal PCC, for example, is important for monitoring environmental inputs to determine the need for a shift in behavioral strategy (57). PCC also receives projections from the medial temporal subsystem and hippocampus that contribute to counterfactual scene construction (58) and from dorsal medial subsystem areas linked to representing others’ mental states, including the superior temporal lobe (STL) and temporoparietal junction (TPJ; 35). We have found previously that connectivity between the posterior hub and STL/TPJ is associated with mentalizing. Exploitativeness is also positively related to connectivity between STL/TPJ and dorsal regions of the posterior core, whereas callousness is negatively related to STL/TPJ connectivity with ventral regions of the posterior core (17). Our experiment diverges fundamentally from typical mentalizing tasks in that it addresses how people learn to cooperate or not based on social feedback. However, the observed relationships with exploitativeness and callousness suggest that social computations in the posterior hub may be an important component of exchanges in which it is advantageous to correctly predict the behavior of a counterpart. It is possible these computations are a necessary, but not sufficient, component of more complex mentalization.

Strengths of this study include a validated reinforcement learning model capturing trial-level behavior and decision-making, a well-characterized clinical sample that allowed for comprehensive assessment of the latent agreeableness–antagonism dimension, a modern neuroimaging protocol with subsecond temporal resolution, and the use of MSEM to demonstrate how neurocomputational signatures of learning can explain associations between stable personality facets and trial-level decision-making. Nonetheless, our observations were contingent on natural variation in personality, precluding strong causal inferences. Likewise, though the robust effects of irrelevant reputation suggest participants perceived the task as social in nature, we cannot definitively rule out that our findings capture a broader learning process that might extend to other, nonsocial bandit tasks as well. Replicating our findings using live social interactions and human counterparts would help resolve this question.

Finally, our sampling strategy recruited participants with a probable or definite diagnosis of BPD in addition to healthy comparison participants. This allowed us to capture the full range of severity for both callousness and exploitativeness, potentially increasing the generalizability of our findings to populations with clinically significant externalizing psychopathology. Reassuringly, while participants in the BPD group were higher than healthy comparison participants on both traits, group status did not attenuate the effects of callousness and exploitativeness on policy PEs or reciprocity in sensitivity analyses. Nonetheless, we cannot definitively rule out the possibility that callousness and exploitativeness function differently in BPD than healthy comparison participants due to interactions with other constructs that are also elevated in BPD. If this were the case, it would limit the generalizability of our findings to more transdiagnostic samples or samples drawn from the general population. It is reassuring that our effects held when we controlled for the traits most likely to be correlated with callousness and exploitativeness in the context of BPD, including negative affect and impulsivity, but the only way to know the true generalizability of our findings is for future studies to attempt replications in the general population and in more clinically diverse samples.

Overall, we found that during social exchanges, cooperation-related learning signals in the default network scaled negatively with callousness but positively with exploitativeness. Furthermore, default network sensitivity to changes in the counterpart’s estimated cooperation rate led to greater social reciprocity in exploitative individuals and lower reciprocity in callous individuals. If the expansion of the default network in anthropogenesis contributed to the unprecedented flexibility in human cooperation, it may have also contributed to the darker side of human sociality.

Methods

All participants enrolled in the study provided written informed consent to participate. All procedures were approved by the University of Pittsburgh Institutional Review Board (IRB Number: STUDY19050210). For more details on participants, procedures, and analyses, refer to our SI Appendix.

Participants.

The final sample (n = 168) included 113 individuals with a probable or definite diagnosis of BPD and 55 healthy comparison participants whose behavioral data passed all quality control checks as described in SI Appendix, section I. Eighteen participants were excluded from imaging analyses due to missing data (n = 1), poor quality imaging data (n = 3), and excessive motion (n = 14; see SI Appendix, section VII). Participants from the two groups were pooled into a single sample to allow for continuous data analysis across a range of symptom severity. Demographic and clinical characteristics of the final sample are presented in SI Appendix, Tables S2 and S3.

Experimental Paradigm.

Participants completed a modified version of a trust game featuring three different trustees, as depicted in Fig. 1B (27). Trustees were introduced prior to the task via a picture of a neutral, White, male face from the NimStim Face Stimulus Set (59) and a short description that included one noteworthy event indicative of the trustee’s “good,” “bad,” or “neutral” reputation.^† Participants were told that they would be playing against the various trustees in the scanner and that the behavior of the trustee would be sensitive to their own choices (even though the reinforcement rate was standardized; see below). Participants made their own determination as to whether the trustees were real players or not. Participants completed a brief questionnaire before and after the task to rate each trustee’s likeability and trustworthiness.

Participants interacted with each trustee for a total of 48 trials. All trustees invested 50% of the time in the first 16 trials, and either 25% (poor) or 88% (rich) of the time in the next 16 trials; the reinforcement schedule reversed in the last 16 trials (i.e., rich-to-poor or poor-to-rich). The order of the trustees was counterbalanced across participants. At the beginning of each trial, participants were told they had $1.00 to either keep or invest with the trustee. If they chose to keep, they would retain the $1.00. If they chose to invest, they could receive either $1.50 if the trustee also chose to invest or be left with nothing if the trustee decided to keep their $1.00. Earnings from each trustee were shown to participants at the end of each trustee block.

Each trial began with a fixation cross, followed by a decision phase, during which participants made their choice to keep or invest (decision phase ended after 2.7 s when no response was made). The participant’s choice was highlighted for 200 ms, after which participants were shown the trustee’s decision for 1.2 s. Critically, participants were shown the trustee’s decision regardless of their own choice, which enabled participants to learn from counterfactual outcomes, with the only difference being that when participants chose to keep, the trustee decision was displayed in gray (as opposed to red or green) to indicate it did not affect earnings for that trial. Intertrial intervals were jittered by combining the unused decision time from the previous decision phase with a base duration sampled from an exponential distribution, with no ITIs allowed to be >1,500 ms or <250 ms.

Factor Analysis.

Subscales assessing the agreeableness–antagonism domain were selected from a battery of self-report instruments, including 1) altruism, sympathy, modesty, cooperation, morality, and trust subscales from the NEO-IPIP-120 (60); 2) attention-seeking, callousness, deceitfulness, manipulativeness, and grandiosity subscales from the Personality Inventory for Diagnostic and Statistical Manual, Fifth Edition (61); 3) exploitativeness, entitlement rage, and exhibitionism subscales from a modified version of the Brief Pathological Narcissism Inventory (62); and 4) assault, verbal hostility, and negativism subscales from the Buss Durkee Hostility Inventory (63). Criteria for selecting subscales from each measure are outlined in SI Appendix, section V.

EFA using maximum likelihood estimation and an oblimin rotation was conducted on the 17 self-report subscales. One- and two-factor solutions were extracted to attain measures of broadband antagonism and its core facets, respectively. Estimated factor scores from the EFA solutions were saved using the tenBerge method for use in subsequent analyses.

Behavioral Analysis.

Multilevel logistic regression was used to examine the effect of trustee reputation, the trustee’s last decision, and the exchange number with the trustee (i.e., trial within trustee) on participant choices to keep or invest. Random slopes were modeled for all three design variables, and a random intercept for subject was included to account for individual differences in cooperation. Age and sex were included as participant-level covariates. These analyses provide an independent corroboration of key behavioral effects relative to the cognitive computational model.

Computational Modeling.

Reinforcement learning models were fit using the variational Bayesian analysis toolbox in MATLAB, which leverages parameter estimates from the full sample to constrain individual parameter estimates, reducing the risk of misestimation in poorly performing participants (64). Models employed a version of Q-learning with a single hidden state for the expected value of the invest action on trial t. The value of the keep action was assumed to be updated reciprocally as detailed in our previous study (27). The expected value of the invest action was updated according to the learning rule:

$$Q(invest{)}_t=Q(invest{)}_{t-1}+\alpha \times [reward-Q{\left(invest\right)}_{t-1}],$$ [1]

where $\alpha$ is the learning rate parameter and $reward-{Q\left(invest\right)}_{t-1}$ is the PE. Choices were modeled using a softmax function that included four free parameters:

$$P(invest{)}_t=\frac{1}{1+\mathit{exp}\left[{\kappa }_{subject} + {𝟙}_{trustee}\left(good\right)\text{ × }{\kappa }_{trustee}\left(good\right) + {𝟙}_{trustee}\left(bad\right) \mathbf{\times }\mathbf{ }{\kappa }_{trustee}\left(bad\right) + \beta \times Q{\left(invest\right)}_t\right]},$$ [2]

where $\beta$ is a temperature parameter reflecting choice stochasticity, and ${\kappa }_{\mathit{subject}}$ is a participant-level parameter that reflects a general bias to invest (vs. keep) regardless of reinforcement history and corresponds to the tendency to invest with the neutral trustee. ${\kappa }_{trustee}\left(good\right)$ and ${\kappa }_{trustee}\left(bad\right)$ represent the shift in the participant’s sharing observed with the good and bad trustees relative to the neutral trustee. These parameters are non-zero only for exchanges with the good and bad trustees, respectively. Four of the five competing models included the same free parameters $(\alpha ,\beta ,{\kappa }_{subject},{\kappa }_{trustee}(good), {\mathrm{a}\mathrm{n}\mathrm{d} \kappa }_{trustee}(bad\left)\right)$ but differed with respect to their payout matrices. A fifth model included trustee-specific learning rates (${\alpha }_{\left(good\right)},{\alpha }_{\left(bad\right)},{\mathrm{a}\mathrm{n}\mathrm{d} \alpha }_{\left(neutral\right)})$ with a single participant-level bias parameter ${(\kappa }_{subject})$ to reflect a general tendency to invest. We have shown in previous work that the models’ parameters are identifiable (27). Critically, the distinguishing feature of the policy model was that the valence of the reward on trials in which the participants correctly predicted a trustee defection was positive relative to trials in which participants incorrectly predicted a defection. See SI Appendix, section VI for the rationale of competing models and corresponding payout matrices. The winning model was selected by comparing the relative evidence for each model using random effects BMC, accounting for the full statistical risk incurred (Bayesian omnibus risk [BOR], 65).

fMRI Processing and Analysis.

Imaging data were acquired in a Siemens Prisma 3T scanner at the Magnetic Resonance Research Center at the University of Pittsburgh. We acquired functional imaging data during the trust task using a simultaneous multislice sequence sensitive to BOLD contrast, TR = 0.6s, TE = 27 ms, flip angle = 45°, multiband acceleration factor = 5, voxel size = 3.1 mm³. We obtained a sagittal MPRAGE T1 scan, voxel size = 1 mm³, TR = 2.3 s, TE = 3.35 ms, GRAPPA 2x acceleration. The anatomical scan was used for coregistration and nonlinear transformation to functional and stereotaxic templates. We also acquired gradient echo fieldmap images (TEs = 4.47 and 6.93 ms) for each subject to quantify and mitigate local inhomogeneity of the magnetic field. Neuroimaging data were preprocessed using FMRIPREP version 20.1.1 (66). See SI Appendix, section VII for full details on preprocessing.

Voxelwise general linear model analyses were conducted using FSL FEAT v6.0 via fmri.factory (67), an R package that streamlines model-based fMRI analyses. Choice and feedback phases of the task were modeled with duration-modulated unit-height boxcar regressors that were convolved with the canonical hemodynamic response function. Participant reaction time on each trial determined the duration of the choice regressor, whereas the duration of the feedback regressor was fixed at 1.10 s. Standardized PE and value signals derived from the policy model were multiplied by the participant’s decision on each trial (−1/1 = keep/invest) to reflect the assumption that policy successes would be associated with greater BOLD activity. Transformed PE and value signals were added as parametric regressors in GLM analyses. PE was aligned to the feedback event and value was aligned to the choice event. To unconfound the influence of response time on BOLD response from the influence of parametric regressors during the choice phase, we convolved the duration-modulated boxcar regressor for the choice event with the HRF, renormalized the peak to 1.0, multiplied the regressor by the corresponding parametric modulator (value), and then summed across trials (cf. 68). Trustee type was included as a within-subjects factor, allowing for linear contrasts among trustees. Nonresponse trials were modeled independently from trials on which participants made responses.

Group-level analyses were carried out using FSL FLAME 1+2 with automatic outlier deweighting (69), which implements Bayesian mixed effects estimation of the group parameter estimates including full Markov Chain Monte Carlo-based estimation for near-threshold voxels (70). Group analyses included age and sex as covariates of no interest. To correct for familywise error at the whole-brain level, we applied the probabilistic threshold-free cluster enhancement method (pTFCE), thresholding whole-brain maps at FWE P < 0.05 (71). This algorithm provides strict control over familywise error and boosts sensitivity to clusters of activated voxels.

To examine how individual differences in default network BOLD response related to learning and behavior, we extracted participant-level regression coefficients (“betas”) from the PE contrast of the whole-brain fMRI GLM analyses based on an a priori specified atlas that combined leading cortical and subcortical parcellations (43, 72, 73). The atlas included 244 (200 cortical + 44 subcortical) parcels assigned to 17 functional networks (74). In total, 37 parcels were assigned to one of the three default network subsystems in ref. 74. All voxels within each DN parcel were included in the extracted betas. To avoid redundancy, fMRI betas were extracted from an intercept-only group model, as age and sex were entered in covariates in all MSEM models. To derive an index of PE modulation of the default network, we entered betas from all parcels assigned to the default network into a factor analysis, extracted a one-factor solution, and saved estimated factor scores using the tenBerge method (75). Likewise, to derive indices of PE signaling in each default network subsystem, we entered betas only from parcels assigned to each subsystem into a factor analysis (three total), extracted a one-factor solution from each, and saved estimated factor scores. The advantage of using factor analyses to derive these indices (as opposed to using a mean composite) is that it allows for individual parcels to be weighted based on their respective contribution to the broader network. Factor scores derived with this method were comparable to those using a more data-driven approach in which all parcels were included and additional factors were extracted to derive subsystem indices (i.e., we extracted 1-, 2-, 3-, and 4-factor solutions using all parcels; see SI Appendix, section IX).

Multilevel Structural Equation Modeling.

Relations between personality, learning signals encoded in the brain, and trial-level behavioral performance were tested using MSEM, which enables the simultaneous examination of multiple predictors and outcomes in hierarchically structured data (76). In MSEM, latent decomposition is used to partition total variance in outcomes and predictors into within- and between-person components. All MSEM models used Bayesian estimation with noninformative priors in Mplus Version 8.7 (77). Bayesian estimation uses all available data and provides similar results to full information maximum likelihood in accounting for missing data (78). For all MSEM models, we report unstandardized and standardized regression coefficients, 95% credible intervals, and Bayesian P values. Bayesian P values are based on the probability of direction test, a hypothesis test that is closely aligned to frequentist null hypothesis significance testing (79). Note, however, that Bayesian posterior probabilities quantify the extent to which the data support a given hypothesis, which provides stronger inference than frequentist approaches that quantify the probability of observing the data under the null hypothesis.

In all MSEM models, random slopes were specified for the effects of trustee reputation, trustee’s previous decision, and trial within trustee on participant decisions to keep or invest (80). To examine the effect of learning signals on trial-level behavior, default network PE factor scores were added as between-subjects moderators of design effects (Fig. 1D). The ability of MSEM to accommodate multiple outcomes allows for formal tests of mediation across levels in the data. In models examining associations between personality, learning signals, and behavior, personality traits were added as between-subjects predictors of default network learning signals. Mediation (the effect of personality on trial-level behavior via cooperation-related learning signals) was tested using a product of coefficients approach. In contrast to conventional mediation analysis, Bayesian mediation analysis constructs credible intervals for indirect effects that are not subject to normality assumptions on the sampling distribution of the estimate of the indirect effect, and which do not rely on large sample approximations, allowing for more exact inferences (81). All models controlled for age and sex as between-subjects covariates.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Anonymized Behavioral, Self-report, Neuroimaging data have been deposited in Open Science (10.17605/OSF.IO/UARHZ) (82).