(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Social learning dynamically shapes moral decision-making by biasing subjective valuation [1]

['Julien Benistant', 'Cnrs', 'Institut Des Sciences Cognitives Marc Jeannerod', 'Bron', 'Valentin Guigon', 'Université Lyon Claude Bernard', 'Lyon', 'Alain Nicolas', 'Genopsy', 'Ch Vinatier']

Date: 2026-07 

Observing immoral behavior increases one’s dishonesty by social influence and learning processes. The neurocomputational mechanisms underlying such moral contagion remain unclear. We tested different mechanistic hypotheses to account for moral contagion. We used model-based fMRI and a new cheating game in which participants were sequentially placed in honest and dishonest social norm contexts. Participants’ cheating behavior increased in the dishonest norm context but was unchanged in the honest one. The best model to account for behavior indicated that participants’ valuation was dynamically biased by learning that others had cheated. At the group level, this valuation bias was not encoded by any specific brain region. Instead, this neural signal depended on individual differences in conformity, and engaged the bilateral lateral prefrontal cortex. During learning, simulation of others’ cheating behavior was encoded in the posterior superior temporal sulcus. Together, these findings provide a mechanistic understanding of how learning about others’ dishonesty biases individuals’ valuation of cheating but does not alter one’s established preferences.

Funding: This research has benefited from the financial support of IDEXLYON from Université de Lyon (project INDEPTH) to JCD within the Programme Investissements d’Avenir (ANR-16-IDEX-0005) and of the LABEX CORTEX (ANR-11-LABX-0042) of Université de Lyon to JCD, within the program Investissements d’Avenir (ANR-11-IDEX-007) operated by the French National Research Agency ( https://anr.fr/ ). This work was also supported by grants from the Agence Nationale pour la Recherche to JCD (ANR-24-NEUC-0001, ANR-24-CE37-4261-01, ANR-24-EXEN-0001), by MITI 2020 CNRS to JCD, by ANR-25-CE41-1847, and by the Initiative of Excellence of the University of Lille (R-CDP-24-006-DePERU) to JB ( https://deperu.univ-lille.fr/ ). The funders had no role in study design, data collection and analysis, decision to publish, or preparation to the manuscript.

We used functional magnetic resonance imaging (fMRI) and a novel paradigm based on a cheating game. This game comprised two types of trials: (i) Solo trials, in which participants could lie about the outcome of a die roll to maximize their earnings, (ii) Predict trials, in which they predicted what another individual, randomly selected from a group of 10, reported in a previous experimental session. The purpose of the Solo trials was to assess the extent to which participants were prone to cheating. For the Predict trials, the goal was to expose participants to the behavior of others and allow them to learn the preferences of a group. Unbeknownst to the participants, the behavior of the group was simulated and controlled so that the participants faced either a dishonest or an honest group of players. With this manipulation, we assessed the effect of social influence in two different contexts. These two contexts have been shown to lead to different levels of conformity because antisocial and dishonest behavior appears to be more contagious than pro-social behavior [ 11 , 36 , 37 ]. The experiment was divided into three blocks. The first was composed of Solo trials and allowed us to estimate participants’ preferences in the absence of social influence. The second and third blocks consisted of interleaved Predict and Solo trials. In each of these two blocks, participants’ predictions concerned either an honest or a dishonest group of participants. This novel design gradually exposed participants to others’ behavior, allowing us to test the fixed versus dynamic accounts of valuation bias and preference shift mechanisms.

In addition to the tested mechanistic accounts, we also considered that inter-individual differences may shape how social information influences cheating behavior. Substantial variability exists in individuals’ tendency to conform to others and in their intrinsic honesty, and prior works show that such inter-individual differences systematically affect how social information is integrated into decision-making [ 6 , 35 ] and also influence how moral trade-offs are resolved [ 16 , 27 ]. We therefore further hypothesized that the dynamic social influence mechanisms captured by the tested computational models would vary across individuals as a function of their propensity to conform, and that their individual parameters would be linked to their brain activity during decision-making.

One brain region that could support such dynamic changes in social influence is the dorsolateral prefrontal cortex (dlPFC), as it represents others’ risk preferences during behavioral contagion [ 6 ]. Moreover, activity in the dlPFC is associated with integration of social norms and moral preferences in decision processes [ 16 , 24 – 26 ] and neuroimaging studies consistently implicate the dlPFC in dishonest choices [ 27 – 29 ]. Regions such as the right temporo-parietal junction (rTPJ) and the posterior superior temporal sulcus (pSTS) are also good candidates to support the dynamic integration of others’ behavior. Indeed, both the rTPJ and pSTS are central to tracking and integrating others’ preferences [ 30 ], and there is evidence that the rTPJ encodes others’ intentions necessary for social choices [ 31 – 34 ].

This figure represents the four different mechanisms of social influence that we considered. We varied whether social influence was due to a shift in one’s preferences (top row of the table) or a bias in the valuation process (bottom row). We then varied whether the process was fixed (left column) or dynamic because it depends on one’s learning about the behavior of others over time (right column) (schematically represented by the top figure, right column). For each model, a formal description is provided based on a utility function (U cheat ) that represents the relative value of cheating for the individual, with two free parameters α and δ. α represents the individual’s sensitivity to monetary reward whereas δ represents the moral cost incurred when cheating. According to the Valuation bias Fixed model, the parameter θ represents the extent of the bias in the participants’ valuation of cheating. According to the Dynamic version, the parameter γ represents the level of conformity of the participant to others’ learned behavior (represented by ). Finally, in the Preference shift models, social influence is represented either by a fixed change in participants’ preferences α and δ (Fixed) or by a dynamic evolution of their preferences based on the learned preferences of others (Dynamic).

Here, we tested different theoretical accounts of the mechanisms underlying how others’ behavior influences one’s own cheating behavior. We considered four social influence mechanisms: a fixed or dynamic valuation bias, and a fixed or dynamic change in individual preferences ( Fig 1 ). First, we assessed which computational model best explains social influence in cheating behavior. Second, we uncovered where in the brain different computational signals of social influence are encoded when we are presented with the opportunity to cheat. Importantly, the different accounts noted above predict that different neural mechanisms underlie social influence on cheating behavior. According to the first two hypotheses (fixed valuation bias and fixed preference shift hypotheses), no behavioral or brain activity changes should occur over time, because they do not consider that learning about the behavior of others would dynamically affect the influence process. In contrast, the dynamic valuation bias and dynamic preference shift hypotheses do predict such changes.

One important limitation of these two previous accounts of social influence is that they do not consider the dynamic aspect of learning. For example, in a new social environment with specific norms (e.g., when one arrives in a new company), one usually lacks knowledge about others’ behavioral tendencies, including whether they lean toward honesty or dishonesty. Consequently, in such new environments, learning through repeated observation of others’ behavior becomes necessary to mitigate uncertainty about others’ true levels of (dis)honesty. Learning social norms in new contexts is particularly important regarding dishonesty because its concealed nature makes accurate observation challenging. Therefore, it is likely that a social learning component might modulate either the valuation bias or the preference shift to adjust how such mechanisms are weighted as the individual learns about others’ honesty levels.

It remains unclear whether social influence shapes individual beliefs, because behavior reflecting social norm compliance is not a perfect proxy for private beliefs. To date, two mechanisms have been proposed to explain how social influence affects one’s choices. The first, called valuation bias, proposes that individuals assign value to social information, which then influences their personal valuation process. This means that simply being exposed to others’ choices alters how individuals perceive the value of the options at stake and potentially influences their own choice [ 3 , 4 , 22 , 23 ]. The second mechanism, referred to as preference shift, proposes that exposure to others’ choices can more fundamentally modify an individual’s own preferences to align with the choices of others. This mechanism implies a more profound change in personal preferences and assumes that individuals become more similar to others in their social environment [ 5 – 8 ]. The two models differ in their assumptions with regard to how the decision-making process is affected. According to the valuation bias model, individuals place value on social information, such as others’ choices, and this additional value biases individuals toward others’ choices [ 3 , 4 , 22 ]. Conversely, in the preference shift model, individuals’ preferences themselves are changed. According to this latter model, individual preferences can be defined as the weights placed on the different components of a choice, such as monetary earnings, consequences for others or the morality of the action. Thus, individual preferences weight each component of the choice when computing decision value [ 22 ]. In this case, it is not the value of social information that leads to social influence but rather a convergence of individuals’ own preferences toward those of others [ 5 – 8 ].

Dishonest behavior, such as cheating, tax evasion, and corruption is pervasive in modern societies. Dishonest behavior can be modified through exposure to others’ immoral behavior [ 1 , 2 ]. For example, observing others behave dishonestly increases one’s likelihood to cheat [ 2 ]. Such influence of others’ behavior on our own choices is not limited to dishonesty but can also be observed in domains such as risk-related decision-making [ 3 – 8 ] as well as pro-social and anti-social behavior [ 9 – 12 ]. Why and how observing others’ dishonest behavior modifies our own moral behavior in non-strategic settings remains an enduring puzzle. A growing literature has highlighted the importance of neurocomputational approaches for understanding how social information, moral preferences and value computations interact to shape human behavior. In particular, recent work has implicated specific brain regions in moral decision-making and described the underlying neural computations [ 13 – 21 ]. Yet, little is known about the neurocomputational mechanisms underlying how social influence modifies moral decisions.

A. The BOLD signal in the bilateral posterior superior temporal sulcus (pSTS), the temporo-parietal junction (TPJ) and the dorsal anterior cingulate cortex (dACC) correlates with the probability associated with participants’ predictions in the Dishonest Group condition at the prediction stage of the Predict trials (p < 0.001 uncorrected and p < 0.05 whole-brain cluster-corrected family-wise error (FWE)). B . Contrast between the brain activity in the Dishonest and Honest Group conditions that correlates with the probability associated with participants’ predictions at the prediction stage of the Predict trials (p < 0.001 uncorrected and p < 0.05 whole-brain cluster-corrected family-wise error (FWE)).

In the second GLM (GLM 2), we added the remaining three computational signals: the probability associated with participants’ predictions at the time of the prediction in the Predict trials, their prediction error at the time of the feedback in the Predict trials and the relative value of the participants’ choice at the choice stage of the Solo trials in every condition (Baseline and Groups; see Methods SI for details). This GLM revealed that the probability associated with participants’ predictions in the Dishonest Group condition is encoded in a large cluster that includes the bilateral posterior Superior Temporal Sulcus and Temporo-Parietal Junction (pSTS-TPJ) and the dorsal anterior cingulate cortex (dACC; cluster’s peak, x,y,z = 57,6,24 and −27,30,27; see Fig 5A and S6 Table ; surviving a FWE-peak correction, see S5A Fig for Betas extracted from an independent ROI). Furthermore, a contrast between the Dishonest and Honest Group condition showed that a part of this cluster activity was significantly greater in the Dishonest Group condition than in the Honest Group condition (cluster’s peak, x,y,z = −18, − 18,3 and 60,6,27; see Fig 5B and S7 Table ; p < 0.05 FWE whole-brain cluster corrected). When looking for regions encoding this computational signal in the Honest Group condition, no voxels survived the cluster-FWE correction (p > 0.05). Additionally, the relative value of the chosen option ( in the Solo trials was negatively correlated with the BOLD signal in the dorsal Anterior Cingulate Cortex (dACC, x,y,z = 18,27,48). However, this cluster did not survive the cluster FWE correction (p(FWE-corrected) = 0.203). Finally, the prediction error at the time of the feedback in the Predict trials, correlated with a large cluster encompassing the bilateral ventral striatum (x,y,z = 18, − 15,24; S5B Fig and S8 Table ).

In the first GLM (GLM 1), we included only the first signal (dynamic valuation bias) as a parametric regressor during the choice phase of the Solo trials in both the Dishonest and Honest Group conditions (see Methods ). The purpose was to assess whether some brain regions encoded this signal in the two conditions. This GLM revealed no significant correlation between brain activity and the dynamic valuation bias in the Solo trials of either Group condition (p > 0.001 uncorrected). One possible explanation for the absence of a group-level effect is that the neural encoding of the dynamic valuation bias depends on individual differences in conformity to others’ behavior. Specifically, participants who behave against the observed norm (e.g., cheating less than the Dishonest Group or more than the Honest Group) may encode this signal differently from participants whose behavior aligns with that of others. Consistent with this interpretation, when focusing on the Dishonest Group condition, a second-level regression analysis revealed that the relationship between the dynamic valuation bias and brain activity varied as a function of participants’ conformity parameter γ. In particular, activity in the left Lateral Prefrontal Cortex (lPFC), the right dorsolateral Prefrontal Cortex (rdlPFC) and bilateral Inferior Parietal Sulcus (IPS) encoded the dynamic valuation bias which was inversely proportional to the conformity parameter, γ, for the Dishonest Group condition (lPFC, x,y,z = −48,42,0; rdlPFC = 30,39,42; IPS = −33, −63,45 and 36, −66,39; Fig 4A and S5 Table ; p < 0.05 FWE whole-brain cluster corrected). To illustrate this effect ( Fig 4B ), individuals’ betas were extracted from 6 mm spheres centered on peak activity in the bilateral lPFC in the Dishonest and Honest Group condition. For negative values of γ, participants show anti-conformity while positive γ values reflect conformity.

We constructed three GLMs to identify how the brain encodes the 4 computational signals of the VB Dynamic model. The first three signals concerned the decision and prediction processes (at the time of choice in the Solo trials and the time of prediction in the Predict trials): (1) the dynamic valuation bias in the Dishonest and Honest Group conditions (choices in the Solo trials); (2) the probability associated with participants’ predictions, when observing Dishonest and Honest group members (predictions in the Predict trials), (3) the relative value of chosen and unchosen options in Solo trials, regardless of conditions (i.e., all blocks averaged together). The fourth signal corresponded to the prediction error at the feedback stage of the Predict trials. Four parametric regressors of no interest were also added: participants’ response times, choice side (left or right), decisions to cheat or not at the choice stage of the Solo trials and their predictions about others’ behavior during the prediction stage of the Predict trials. All the parametric regressors were orthogonalized due to a strong correlation between the parametric regressors of interest and those of no interest. Finally, we added the participants’ degree of conformity γ for both the Dishonest and Honest Group conditions as a second-level covariate.

Finally, we also examined pairwise correlations between fitted parameters ( S4E Fig ). This analysis revealed a negative correlation between β and δ, suggesting that participants with stronger moral costs tended to show less stochasticity in their choices. We also observed a negative correlation between γ D and γ H , indicating that participants who conformed more strongly to the Dishonest Group tended to conform less to the Honest Group. We also ran a robustness check. As a final robustness check, we examined whether participants’ explicit reports of changes in moral beliefs about cheating affected the behavioral and computational results. Based on the debriefing questionnaire, 13 participants reported a change in their beliefs about the morality of cheating in the task, whereas 18 did not. Additional regression analyses showed no significant differences between these two groups in cheating behavior, prediction accuracy, or predicted cheating behavior ( S4 Table ). Moreover, behavior in both groups was best explained by the VB Dynamic model (pEP > 0.98 for both groups), and the fitted parameter values did not differ significantly between groups (Wilcoxon rank-sum tests, p > 0.05 for all parameters). Thus, the main behavioral and computational results were not driven by participants who reported a change in explicit moral beliefs about cheating.

Formally, the learned α parameters are significantly higher for the Dishonest Group than the Honest Group (mean α, Dishonest group: 0.437, Honest Group: 0.090, two-sided rank sum test, z = 4.78, p < 0.001; S4C Fig ). We also found that both learned α parameters are significantly positive (Wilcoxon signed rank test, z(Dishonest, Honest) = 4.86, p < 0.001). Conversely, the learned moral cost δ parameters are significantly lower for the Honest Group than the Dishonest Group (mean δ, Dishonest group: −0.089, Honest Group: −1.736; two-sided rank sum test, z = 6.46, p < 0.001; S4D Fig ). Moreover, the learned moral cost δ parameter is significantly different from zero in the Honest Group condition (Wilcoxon signed rank test, z = −4.86, p < 0.001) but not in the Dishonest Group condition (Wilcoxon signed rank test, z = −1.10, p = 0.272). This implies that participants recognized that members of the Dishonest Group were more likely to cheat than members of the Honest Group for a given set of payoffs. However, the variance parameter of the distribution for the learned parameter α is significantly higher for the Honest than the Dishonest Group. Conversely, the variance parameter of the distribution for the learned parameter δ is significantly higher for the Dishonest Group than the Honest one (mean var. α, Dishonest group: 0.396, Honest Group: 0.855; mean var. δ, Dishonest Group: 0.531; Honest Group: 0.312; two-sided rank sum tests, z = −4.41, p < 0.001 and z = 3.18, p = 0.001, respectively). Furthermore, the difference in variance for the parameter δ is significantly lower than the difference for the parameter α (α diff: M = −0.459, δ diff: M = 0.219, two-sided rank sum test, z = −5.86, p < 0.001). This finding parallels the behavioral results which indicate that participants’ prediction accuracy is lower in the Honest Group condition than in the Dishonest Group condition.

A. Negative relationship between the BOLD signal in the left lateral prefrontal cortex (lPFC) and the dynamic valuation bias signal (i.e., the learned probability that the other would have cheated), as a function of participants’ conformity parameter γ, at the decision stage of the Solo trials of the Dishonest Group condition (p < 0.001 uncorrected and p < 0.05 whole-brain cluster-corrected family-wise error (FWE)). B. Individuals’ beta extracted from 6 mm sphere centered on peak activity in the left lPFC (right dlPFC) are negatively correlated with their conformity parameter γ in the Dishonest Group condition.

Third, concerning the learning of others’ preferences, we found that the predicted probability that others would cheat, estimated by participants, converged towards the average simulated probability to cheat of each group ( Fig 3C ). We also found that the dynamic valuation bias converged towards the mean probability to cheat (to be honest) for the Dishonest (Honest) Group condition ( Fig 4C ).

Second, we observed that the conformity parameter γ is significantly higher in the Dishonest Group condition than in the Honest Group condition (Two-sided rank sum test, z = 3.45, p < 0.001; Figs 3B and S4 B ). Additionally, the parameter γ is significantly different from zero in the Dishonest Group condition (mean: 0.473 ± 0.106, Wilcoxon signed rank test, z = 3.45, p < 0.001, S4B Fig ) but not in the Honest Group condition (mean: −0.013 ± 0.094, Wilcoxon signed rank test, z = 0.02, p = 0.984, S4B Fig ). This indicated no reliable conformity effect in the Honest Group condition while there is a clear conformity effect in the Dishonest Group condition. Indeed, higher γ values indicated a greater tendency to align with others’ behavior (more cheating in the Dishonest Group condition and less cheating in the Honest Group condition, see S5A Fig for a simulation). Furthermore, we examined whether an analog of γ could be observed in the raw data by correlating the fitted γ parameters with participants’ change in cheating relative to baseline (mean cheating in each Group condition minus mean cheating in the baseline). These correlations were in the expected direction, positive in the Dishonest Group condition and negative in the Honest Group condition, although the association was weak and only significant in the Honest Group condition ( S5B Fig ). Thus, aggregate changes in cheating provide a coarse behavioral proxy for γ, but should not be interpreted as a direct equivalent of the fitted conformity parameter. This is because such aggregate measures do not capture the dynamic aspect of conformity in our experiment. Finally, one could argue that participants’ conformity may be explained by their own preferences or by the accuracy of their predictions. Yet, these alternative hypotheses are not supported by additional analyses which show that γ is neither correlated with the participants’ own moral cost δ, nor with participants’ average prediction accuracy ( S5C – S5E Fig ). However, we cannot completely rule out these alternative explanations with the present experimental design.

We then explored the estimated parameters of the winning VB Dynamic model. First, the parameter α (mean: 0.235 ± 0.021, Wilcoxon signed rank test, z = 4.86, p < 0.001, S4A Fig ) was significantly positive, as well as the inverse temperature parameters β (mean: 2.821 ± 0.293, Wilcoxon signed rank test, z = 4.86, p < 0.001, S4A Fig ) and β Oth (mean: 1.100 ± 0.062, Wilcoxon signed rank test, z = 4.86, p < 0.001, S4A Fig ). However, these parameters are constrained in our model to be significantly greater than zero. Conversely, δ is negative and significantly different from zero (mean: −0.727 ± 0.190, Wilcoxon signed rank test, z = −3.41, p = 0.001, S4A Fig ). Overall, this was consistent with participants’ likelihood of cheating increasing with the relative earnings of cheating, while being limited by a fixed moral cost associated with cheating.

A . Protected Exceedance Probability (pEP) of the different candidate models. The higher the pEP, the more likely a given model explains the group’s behavior relative to the others. PS, Preferences shift; VB, Valuation Bias. The dashed line indicates chance level. B . Difference between the value of the conformity parameter γ in the Dishonest Group condition and the conformity parameter γ in the Honest Group condition for each participant. The stars represent the result of a two-sided rank-sum test between the two conditions. C. Mean inferred probability that others cheated during Predict trials in the Dishonest Group condition (red) and the Honest Group condition (blue). Shaded areas are standard errors. The red (blue) dashed line corresponds to the mean proportion of cheating of the Dishonest (Honest) group members. D . Mean dynamic valuation bias over Solo trials in the Dishonest Group condition (red, probability that others would have cheated) and in the Honest Group condition (blue, probability that others would have been honest). Shaded areas are standard errors. The red (blue) dashed line corresponds to the mean proportion of cheating (honest behavior) of the Dishonest (Honest) group members. *** p < 0.001. The data underlying the figure can be found in the Figures folder of the OSF repository.

For the fixed hypothesis, we evaluated two models. The first posits that social influence arises from a shift in individuals’ preferences (Preferences Shift, PS Fixed ) [ 5 , 6 , 12 , 42 ]. The second suggests that social influence results from a bias in the valuation process (Valuation Bias, VB Fixed ) [ 3 , 4 , 22 ]. For the dynamic hypothesis, we examined versions of these models in which influence correlated with what participants learned of others’ cheating tendencies across trials. Thus, in the PS Dynamic model, preference change was hypothesized to result from a weighted average between participants’ own preferences and the inferred preferences of others at a given time (PS Dynamic model). This weighted average is governed by two distinct free parameters, and , for each parameter of the others’ utility function ( for the parameter α and for the parameter δ). They are specific to each group condition, indicating the participants’ levels of conformity. For the VB Dynamic model, we considered that the decision value in participants’ choice models is modulated by the probability that others would have cheated (or been honest) in the Dishonest Group (Honest Group) condition during a particular Solo trial, as calculated by the BPL Self model (VB Dynamic model). A parameter, , adjusts the probability and represents the extent of participants’ conformity to others’ behavior. This parameter was individually estimated for each participant in each group condition.

Finally, we proceeded to the selection of the model that best explained social influence in our experiment. We explored two hypotheses, each delineating social influence in distinct ways: one as a fixed phenomenon, independent of participants’ learning about others’ cheating behavior; and the other as a dynamic and learned phenomenon, interconnected with participants’ understanding of others’ conduct.

Next, we evaluated which model best explained participants’ social learning behavior during the Predict trials in the Dishonest and Honest Group conditions. We tested various candidate models derived from the Bayesian Preference Learning (BPL) model [ 41 ]. These models assume that participants start with normally distributed priors regarding others’ preferences, which allows participants to infer the likelihood of cheating when they predict others’ choices during the Predict trials. Post-feedback, a prediction error, that compares their prediction with the actual behavior, refines the posterior distribution of others’ preferences. Within the BPL model, we considered three types of priors, each representing distinct learning biases. The first involved participants using their own preferences as priors for both conditions, a process termed self-projection (BPL-Self). The second utilized participants’ priors based on their own preferences for the initial group condition and the learned preferences from the preceding group for the subsequent group condition (BPL-Others). This captures the persistence of prior learning. Finally, the third model combined aspects of the previous two types of priors. When learning about the second group’s behavior, participants combined their own preferences with the learned preferences regarding the previous group (BPL-SelfOthers). Furthermore, we investigated whether participants inferred others’ preferences according to the utility functions previously mentioned. Specifically, participants could believe that others’ decisions aligned with the utility function used to simulate others’ behavior (Variable cost of cheating, [ 29 ]), or they might consider the utility function they used for their own decisions (Fixed cost of cheating) ([ 39 , 40 ]; see Methods ). We found that the BPL model with self-projection (BPL-Self) and the utility function with a fixed moral cost of cheating was the best to explain our participants’ predictions (pEP: 0.986, S1C Fig ).

First, we used the participants’ decision to cheat (i.e., choose the dishonest option) across all three blocks. Previous work shows that honesty-based choices can be described by two utility functions. A first utility function (Fixed cost) assumes that individuals endure a fixed moral cost when they are cheating [ 39 , 40 ]. A second utility function (Variable cost) assumes that the cost of being dishonest is dependent on the absolute gains associated with it [ 29 ] (see Methods SI). We considered extreme cases in which participants always used just one of these two utility functions (Fixed cost only or Variable cost only) and other cases in which they used one utility function in some blocks and the other in different blocks (Mixed variants). Results of the model selection showed that the utility function with a fixed moral cost of cheating best explained participants’ cheating behavior across all blocks (protected exceedance probability (pEP) = 0.999, S1B Fig ).

Our computational approach to analyze social influence encompassed two components: a social learning component and a social influence component. The social learning component accounts for how participants predicted others’ cheating behavior and updated their own beliefs about others’ preferences. The social influence component reflects how participants cheated (or not), based on their own preferences and based upon how they were influenced by what they learned about others’ cheating behavior. To simplify the selection process between our candidate models, we used a three-step selection procedure based on a group-level random-effect Bayesian model selection [ 38 ]. We started by defining which utility function, among the fixed moral cost and the variable moral cost, best described how participants chose between the honest and the dishonest options. Subsequently, we determined the best-fitting learning model that explained participants’ predictions in blocks 2 and 3. Finally, we conducted model selection among the four candidate models that represent the different social influence mechanisms. Below, we describe each of these steps in more detail.

We next examined whether the order of exposure affected cheating behavior or prediction accuracy. We included a regressor that indicates whether participants were first exposed to the Dishonest or the Honest Group condition. In the model 1 ( S2 Table ), this regressor was not significant, indicating that cheating behavior was not affected by the order of presentation (Margins Dishonest Group condition first = −0.067 ± 0.100, z = −0.67, p = 0.503). Furthermore, in the model 2 ( S2 Table ), this regressor was negative and significant showing that prediction accuracy was higher overall when participants began with the Dishonest Group condition (Margins Dishonest Group condition first = −0.045 ± 0.018 z = −2.52, p = 0.012). This might be due to the greater difficulty to predict the others’ cheating behavior in the Honest Group condition, which may have produced a spillover effect on the participants’ ability to predict the behavior of the Dishonest Group condition.

To ensure that the observed differences were attributable to the groups’ actual behavior, and not differences between their learning performance with respect to the Honest and the Dishonest group’s behavior, we first tested whether participants’ prediction accuracy exceeded chance. Participants’ accuracy was indeed significantly above the chance level of 50% in both conditions (Wilcoxon signed rank tests, z = 4.87 and 4.49 for Dishonest and Honest Group conditions, respectively, p < 0.001 for both groups; Fig 2D ). Moreover, Fig 2E shows that participants’ predictions of others’ cheating frequencies closely aligned with the groups’ mean cheating behavior in both conditions. Consistent with this, participants predicted cheating significantly more often in the Dishonest Group condition than in the Honest Group condition (Margins Dishonest > Honest: 0.540 ± 0.033, z = 16.2, p < 0.001; see model 3 in S2 Table ). However, participants were also more accurate when predicting the Dishonest group’s behavior than that of the Honest group (Margins Dishonest > Honest: 0.179 ± 0.022, z = 8.20, p < 0.001; see model 2 in S2 Table ). Therefore, we tested whether differences in prediction accuracy could account for the observed difference in cheating behavior between the Dishonest and Honest Group conditions. Using the same regression structure as model 1 in S2 Table but adding participants’ average prediction accuracy in each Group condition as a covariate, we found that including accuracy did not alter the effect of group condition on cheating probability (model 1, S3 Table ). A second regression that controlled for each participant’s difference in prediction accuracy between conditions yielded the same result, i.e., the difference in cheating behavior across conditions remained unchanged (model 2, S3 Table ).

Participants were significantly more likely to cheat in the Dishonest Group condition than in any other condition ( Fig 2C , margins contrast Dishonest > Baseline: 0.148 ± 0.037, z = 4.04, p < 0.001, Bonferroni correction for 3 comparisons; margins contrast Dishonest > Honest: 0.101 ± 0.027, z = 3.72, p = 0.001, Bonferroni correction for 3 comparisons; see model 1 in S2 Table ). No significant difference in cheating behavior was observed between the Honest Group condition and the Baseline (Margins Honest > Baseline: 0.047 ± 0.025, z = 1.90, p = 0.170, Bonferroni correction for 3 comparisons).

In the Baseline condition, participants cheated in 55.3% of the Solo trials. They cheated in 69.6% of Solo trials in the Dishonest Group condition and in 60% of the Solo trials in the Honest Group condition ( Fig 2C ). To test the differences between each condition, we used a mixed-effect logistic regression on whether the participants cheated or not in a given trial with the following independent variables: a categorical variable coding for the condition (1: Baseline, 2: Dishonest Group and 3: Honest Group); a binary variable coding for the order of presentation of the Dishonest Group condition (1: first and 0: second); a variable coding for the trial number within each block; a variable coding for the difference between the payoff of cheating and honesty; and, a variable coding for the absolute difference between the die value of the cheating option and the honest one. We also added demographic variables such as the participants’ sex, age, and occupation (i.e., student or not). All the regressions reported in this paper used the same independent variables except when indicated otherwise.

From the initial pool of 32 participants, one was excluded from all analyses due to constant cheating and disbelief that they were predicting real people’s decisions. Additionally, three more participants were excluded from the model-based analyses because they always cheated or always were honest in the Baseline Solo trials (see Methods ). A translated version of the participants’ instructions and debriefing questionnaire is available in SI.

The experiment comprised three blocks: a Baseline block of 50 Solo trials to discern individual preferences in the absence of social influence, and two subsequent blocks, each consisting of 50 Solo and 50 Predict trials presented in an interleaved manner ( Fig 2B ). This sequential arrangement aimed to gradually expose participants to others’ behavior, to test how learning about the others’ (dis)honesty influenced participants’ own decisions in the Solo trials. In these two blocks one block featured a dishonest group (Dishonest Group condition), whereas the other featured an honest group (Honest Group condition, see Methods and S1 Table ). The presentation order of the latter two blocks was randomized across participants.

The “Predict” trials mirrored the Solo trials but required participants to predict whether another individual would report their die roll accurately or not under similar circumstances. Participants were informed that, on each trial, an individual was randomly selected from a pool of 10 anonymous people who had previously completed the task. In reality, the group’s behavior was computer-simulated, which allowed controlled manipulation of their cheating behavior. Participants’ predictions were highlighted in red during a jittered ‘confirmation’ stage. Finally, at the end of each Predict trial, feedback was provided to the participant by highlighting the correct response in green. Accurate predictions were rewarded, whereas inaccurate predictions were not (see Methods for details).

A. The experiment consisted of 2 types of trials. In the Solo trials, participants had to report the outcome of a die roll among two choices. An honest report was always less rewarded than a dishonest one, which made cheating the more profitable choice. The possible value of cheating on each trial varied, and ranged from €2 to €8 in increments of €2 between the honest (low payoff) and dishonest (high payoff) choices. The brute value of honest choices varied from 0 to 8 Euros and that for dishonest choices from 2 to 10 Euros. The Solo trials ended with a confirmation step in which the participants’ screen froze without confirmation of her choice, to reinforce the feeling that the decision was anonymous. In the Predict trials, participants had to predict whether another individual, randomly selected from a group of 10, cheated or reported honestly. Participants’ response was highlighted in red during a jittered confirmation step. Finally, participants received feedback about their prediction. The correct response was highlighted in green. B. Our experiment was divided into 3 blocks. The first one, the Baseline, consisted of 50 Solo trials and allowed us to assess the participants’ preferences. The two following blocks were a mixture of 50 Predict and 50 Solo trials that were interleaved. This structure allowed us to expose participants to others’ behavior and to assess the effect on participants’ decisions. In each of these two blocks, participants’ predictions concerned either an honest or a dishonest group of participants. The order of presentation of the Honest and Dishonest Groups was randomized between participants as either the second or third blocks. C . The mean proportion of cheating in the Dishonest Group condition was significantly higher than in the two other conditions. The statistical analysis was a pairwise comparison with Bonferroni correction extracted from a mixed-effect logistic regression (see model 1 in S2 Table for details). D . Mean prediction accuracy in both the Dishonest Group and the Honest Group conditions. The black line indicates chance level. Stars within bars indicate the results of a one-sided signed rank test comparing the prediction accuracy with the chance level. The other statistical analysis comes from a mixed-effect logistic regression (see model 2 in S2 Table for details). E . The mean of predictions that the other cheated over Predict trials for both the Dishonest and Honest Group conditions. The red (blue) dashed line corresponds to the mean proportion of cheating of the Dishonest (Honest) group members. In all figures, bars or shaded areas correspond to standard errors. *** p < 0.001, ** p < 0.01, * p < 0.05. The data underlying the figure can be found in the Figures folder of the OSF repository.

We scanned 32 participants using fMRI while they played the cheating game, which included two types of trials: Solo and Predict ( Fig 2A ). Solo trials required participants to observe a die roll result (‘stimulus’). Then, participants were presented with two dice (‘choices’), one presenting the original die roll result and the other showing an inaccurate die roll. Each die was associated with a different payoff. Notably, the accurate die always offered a lower payoff than the inaccurate one, which presented participants with a moral dilemma between honesty and maximization of their earnings. To modulate the difficulty of these dilemmas, the payoffs varied across trials. Participants had as much time as they wanted to choose which die roll to report. Once their choice was made, the screen froze for 500 ms (‘confirmation’). No visual feedback indicating their choice was provided to reinforce the feeling that the decision was made privately (see Methods for details).

Discussion

Individuals possess only a limited understanding of others’ opinions, intentions and preferences. Yet, social influence hinges on the ability to infer others’ likely actions. Our findings indicate that when we observe others’ cheating behavior, our brain monitors their inclinations toward dishonesty, not by shaping our own moral preferences, but by biasing our assessment of the value of cheating (valuation bias). Moreover, this valuation bias is dynamic. It reflects progressive learning about the moral inclinations of others. At the brain level, participants conformity modulates how the dynamic valuation bias is encoded in the left lPFC and right dLPFC when observing dishonest behavior. Moreover, others’ dishonesty level is encoded in the TPJ-pSTS when participants predict their behavior. These results provide a dynamic and mechanistic understanding of social influence in cheating, and demonstrate a previously uncharacterized interplay between social learning and social influence processes.

At the behavioral level, participants changed their cheating behavior when they were exposed to the Dishonest Group, but not the Honest Group. One possible explanation for this asymmetry is that participants learned more accurately about the Dishonest Group than the Honest Group. Consistent with this idea, prediction accuracy was indeed higher in the Dishonest condition. However, in regression analyses, when we controlled for differences in prediction accuracy, the effect of group condition on cheating behavior remained unchanged. Furthermore, the conformity parameter estimated by the VB Dynamic model was significantly larger when participants observed the Dishonest Group than the Honest Group. This indicates a stronger tendency to align behavior with dishonest norms. Importantly, these conformity parameters did not correlate with participants’ mean prediction accuracy. These additional analyses show that, while learning accuracy differs between conditions, it cannot account for the asymmetric impact of social influence. Overall, these results are consistent with research showing that anti-social behaviors such as cheating are more contagious than pro-social behavior or honesty [11,36,37]. In addition, social proximity exacerbates this asymmetry [11] and can improve social learning [43]. However, in our experiment, group members were anonymous. Thus, the asymmetry of influence cannot be explained by different levels of social proximity between the participants and others.

In our experimental settings, participants’ cheating decisions were observable to the experimenters at the individual level. Previous studies have shown that such observability can influence behavior, even when participants are informed that their decisions are anonymous [40,44]. This limitation is inherent to fMRI-based, model-driven approaches, which require trial-by-trial observation of individual decisions. To mitigate potential concerns, participants were clearly instructed that their choices were anonymous and that data analysis would be conducted by an experimenter different from the one present during the scanning session, under strict anonymity (see instructions in SI). Importantly, the degree of observability was held constant across all experimental conditions. Moreover, recent evidence suggests that even higher levels of observability than those present in the current study do not substantially alter the influence of empirical social norms on dishonest behavior [45]. It is therefore unlikely that the observability of participants’ decisions can account for the differences observed between conditions in our experiment.

Two main mechanisms of social influence have been proposed previously: an adjustment in individuals’ preferences [6,12,23] or a bias in the individuals’ valuation process [3,4]. Previous work proposed that these two mechanisms are dependent on the accessibility of social information. Specifically, when individuals have full access to others’ behavior or beliefs, preference shifts may be more likely to occur [22]. In the present study, however, we find that a dynamic valuation bias provides a better account of social influence even when others’ behavior is fully accessible and accurately learned. Our model assumes that participants dynamically infer the preferences of others over time (social learning component). This internal representation subsequently biases participants’ valuation process when confronted to the opportunity to cheat (social influence component). This knowledge leads to a change in the value people attribute to the different options. Rather than implying blind imitation or a convergence of preferences, this mechanism is consistent with the view that social influence involves constructing a causal understanding of others’ behavior, which then informs individual decision-making [33,46,47].

One original aspect of our experiment is the gradual exposure of participants to others’ behavior. Past experiments have either presented others’ choices in a block of trials or did not consider the learning dynamics [4,6]. In our task, participants learned relatively quickly, as they were able to predict others’ behavior accurately after approximately 10 trials. While this rapid learning may limit the range over which learning dynamics can be observed, our dynamic model of social influence nevertheless provided a better account of participants’ behavior than fixed models, which suggests that static approaches may not fully capture the processes underlying social influence. A natural extension of this work would be to examine more volatile or uncertain social environments, which would be expected to slow learning and further differentiate dynamic from fixed mechanisms [48].

Our findings also shed light on the neurocomputational mechanisms that support social influence. When participants were exposed to a Dishonest Group, we observed differences in dynamic valuation bias-related activity in the right dlPFC and left lPFC as a function of individual conformity levels. Specifically, greater valuation bias-related activity in these regions was associated with lower conformity, which reflects increased resistance to the observed cheating social norm of the dishonest group. Notably, we did not observe a main effect of the dynamic valuation bias at the group level, which indicates that this neural signal is not uniformly expressed across individuals but rather depends on individual differences in conformity.

This neural pattern mirrors the behavioral results which show that larger valuation biases were associated with a higher likelihood to cheat. The observation that the lPFC dynamically tracks valuation bias extends previous work showing that this region encodes others’ perspectives on choice behavior, whether moral or non-moral [6,16], and plays a role in moral and ethical decision-making through processes of self-regulation, integration of moral values [14,16,28,49–51], and norm compliance [26,52]. Importantly, prior studies examined contexts in which social norms were explicitly known from the outset, rather than learned over time as in the present experiment. Here, we show, additionally, that the lPFC tracks the evolution of the social norm (i.e., the behavior of others) depending on the participants’ degree of conformity. Together, these findings suggest that the lPFC dynamically monitors the social environment and supports the integration of social information into the valuation processes that guide decision-making.

In our BPL model, based on the model developed by [41], participants start with priors about the preferences of the others, which are then used to predict their likelihood to cheat. A prediction error updates those priors and corresponds to the difference between the predicted likelihood to cheat and the other’s actual behavior. During the feedback phase, this prediction error was encoded in the bilateral ventral striatum. This region has been demonstrated to encode prediction errors related to social norms, a process required for learning about the behavior of other individuals [53,54].

When examining brain regions involved in representing beliefs about the dishonesty of group members, we found that activity in the bilateral pSTS and TPJ tracked the probability associated with participants’ predictions during the prediction phase. Notably, this effect was present in the Dishonest Group condition but not in the Honest Group condition. This pattern is consistent with the behavioral asymmetry in learning observed between conditions and aligns with previous findings implicating these regions in simulating others’ learning and representing others’ preferences [30,55]. These results suggest that the pSTS–TPJ supports the computation of signals required for learning and predicting others’ behavioral tendencies, particularly in contexts in which participants show conformity to dishonest behavior.

Importantly, we found no evidence that the pSTS–TPJ directly mediates the social influence process itself, as the dynamic valuation bias was encoded in the bilateral lPFC. In contrast, prior studies reporting pSTS–TPJ involvement in decision-making typically examined situations in which social information was directly relevant for choice, such as cooperation [56], consensus formation [30] or when revising one’s choice [57]. In the present study, participants did not require social information to decide whether to cheat; rather, their decisions could be indirectly influenced by what they learned about others’ behavior. Consistent with this interpretation, previous work on social influence has not observed pSTS–TPJ engagement when social learning was absent [4,6].

The behavior of the individuals used to generate the group decisions was simulated using a utility model that included a variable cost of cheating, whereas the BPL model that best described participants’ own decisions and learning relied on a model which included a fixed cost of cheating. In principle, such a mismatch could make learning about others’ cheating behavior more difficult, as participants would be inferring patterns generated by a model different from their own. However, our behavioral results suggest that participants were nevertheless able to form accurate beliefs. Their predictions were significantly better than chance and converged toward the true mean cheating level in both Group conditions. Several factors may explain the discrepancy between the generative model of the pilot participants and the model that best fits our main sample. First, the pilot participants completed the task in a different context (outside the MRI scanner), which may have influenced their decision strategies. Second, unobserved demographic or psychological differences between pilot and main-sample participants could also contribute to differences in model parameters. Nonetheless, our analyses indicate that participants’ ability to learn about others’ behavior was not impaired in a way that would account for the observed asymmetry in conformity.

A limitation of our study is that the parameter-recovery analyses do not perfectly support the validity of our model-fitting procedure (S3A Fig). This likely reflects the fact that our experimental design was not originally optimized for parameter identifiability across multiple complex components, namely social learning and social influence, which are estimated together in the VB Dynamic model. Specifically, our parameter recovery was weak for one parameter, the others’ temperature parameter (β Oth ) and limited for two parameters (α and β). However, the main parameters of interest in the VB Dynamic model, i.e., the conformity parameters γH and γD, were relatively well-recovered. The limited recoverability of the temperature parameter may reflect, at least in part, the rapid learning observed in participants. Faster learning likely reduced variability in the inferred relative value of cheating across Predict trials, thereby constraining the information available to estimate parameters related to decision stochasticity. Given that the temperature parameter primarily modulates variability in prediction behavior during social learning, this reduction in variability may have contributed to its poor recovery. Thus, these model-based results should be interpreted with caution. However, they constitute a first step toward a better integration of social learning into social influence processes. Future work may build on the present framework using experimental designs tailored to improve parameter identifiability, by increasing environmental volatility or uncertainty to slow learning dynamics.

Overall, our model explains how individuals generate a formal representation of others’ preferences which then biases their valuation process. We report that the pSTS-TPJ and bilateral lPFC compute two separate signals: a social learning signal in the pSTS-TPJ and a social influence signal in the bilateral lPFC. These computational roles elucidate the dynamic link between social learning and social influence in the context of cheating behavior. The representation of others’ behavioral tendencies is encoded in the pSTS-TPJ region during learning, whereas the dynamic bias engages the left lPFC and right dlPFC. These findings have implications for developing new computational theories of social influence that could be useful in different contexts. For example, a recent fMRI study reported that cognitive processes underlying social understanding were more aligned after natural consensus-building conversation, often used for social influence [58]. Much research on social influence has focused on public compliance, setting aside the long-lasting effects of social interaction on private cognition [59,60]. In fact, consensus-building conversation not only aligns neural responses within groups (reflecting social compliance), but this alignment can generalize to novel stimuli that were not discussed [58]. This emphasizes an important role of subjective valuation of others’ learned preferences (an effect explained here by the valuation bias) in understanding social influence. More broadly, our study highlights how social learning dynamically shapes moral decision-making by biasing subjective valuation processes rather than altering stable moral preferences. Such mechanisms may constitute a general computational principle through which social environments shape individual decision-making across domains. Understanding these dynamic social influence processes may be critical for explaining how dishonest or honest norms propagate within social networks.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3003889

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/