Introduction

Charities play vital roles in addressing societal challenges, but depend heavily on individual donations to sustain their activities. Understanding what motivates people to give is therefore critical for supporting charitable efforts. At the same time, individuals differ markedly in their willingness to donate (Auten & Rudney, 1990; Marsh et al., 2014; Nakamura et al., 2025). Identifying the sources of these differences is key to promoting prosocial behaviors at scale, strengthening social cohesion, and enhancing well-being (Aknin et al., 2013; Vanags et al., 2025).

Substantial evidence shows that individuals’ choices and decisions are shaped by those of others, a phenomenon referred to as social influence (Cialdini & Goldstein, 2004; Goldstein et al., 2008; Schultner et al., 2024; Toelch & Dolan, 2015; S. Zhang et al., 2025). In the domain of charitable giving, both laboratory and field studies demonstrate that individual generosity is influenced by others’ prosocial behavior (Chierchia et al., 2020; Nook et al., 2016; Shang & Croson, 2009). For instance, presenting donors with information about a large prior contribution increased individual donations in a public radio campaign (Shang & Croson, 2009), and exposure to higher average donations from others similarly elevated giving (Nook et al., 2016). These findings reveal that others’ choices are a powerful driver of charitable giving.

However, most prior studies focused on how people adjust prosocial behavior in response to a single individual’s decision (Molleman et al., 2022; Shang & Croson, 2009), whereas studies involving multiple others typically summarize social information using an aggregate cue such as the group average or majority level (Nook et al., 2016; Panizza et al., 2021). People now encounter social information from diverse sources in increasingly interconnected environments. In real-world fundraising, such information is often presented repeatedly and derived from multiple donors, reflecting the dynamic distribution of charitable giving beyond single-source influence. Understanding how dynamic sequences of social information shape individual giving is therefore critical to identify effective ways to promote prosocial behaviors in modern society. Yet, despite the widespread availability of diverse social information, how sequential exposure to donation information from multiple donors shapes individual giving remains poorly understood. In particular, it is unclear how the average level and variability of others’ donations together influence charitable decisions.

In recent years, reinforcement learning (RL) models, characterizing direct learning from the environment, have been increasingly applied to the study of social decision-making (da Silva Pinho et al., 2024; Lockwood & Klein-Flügge, 2021; Olsson et al., 2020). These models help elucidate the dynamic processes underlying learning in social contexts (Frolichs et al., 2022; Jin et al., 2023; Joiner et al., 2017), and their extensions further capture how humans learn from others by incorporating social information into value updating and decision-making (Burke et al., 2010; Charpentier et al., 2024; Najar et al., 2020; L. Zhang & Gläscher, 2020; Zhou et al., 2024). Such approaches enable the quantification of latent cognitive processes underlying social influence. In the current study, we employed RL-based modeling to disentangle the computational mechanisms underlying changes in donation behavior in the face of diverse donations from others. Importantly, these models can derive individual-specific computational phenotypes that capture mechanistic variability associated with personality and psychopathology-related traits (Liu et al., 2025; Patzelt et al., 2018; Rhoads et al., 2024).

Social information shapes charitable giving, yet individuals vary considerably in the extent to which they incorporate others’ generosity into their own decisions (Chierchia et al., 2020; Díaz-Gutiérrez et al., 2024; Lee & Chung, 2022). Such individual difference is a core feature of social functioning and offers a valuable lens for understanding how people respond to others’ behavior in prosocial contexts. However, little is known about the determinants of these individual differences - namely, why some individuals substantially adjust their giving toward observed others, whereas others are less responsive. Identifying the sources of heterogeneity in social information use helps clarify when and for whom social information promotes adaptive prosocial behavior, thereby informing targeted strategies to enhance prosocial behaviors (Tump et al., 2024).

Psychopathy and empathy may influence how individuals use social information in charitable contexts. Psychopathy is a personality construct characterized by impulsivity, superficial charm, being callous and antisocial behaviors (Hare, 2006). Its core traits are increasingly recognized as varying across the general population (Guay et al., 2007; Hare & Neumann, 2008). Psychopathy has been associated with reduced prosocial behaviors (Contreras-Huerta et al., 2022; Rilling et al., 2007), diminished sensitivity to outcomes affecting others (Cutler et al., 2021; Rhoads et al., 2025), and a heightened belief in environmental volatility (Atanassova et al., 2025). In contrast, empathy refers to the capacity to understand and resonate with the affective experiences of other people (Singer & Lamm, 2009), and is linked to greater prosocial behaviors (Lockwood et al., 2014), enhanced learning in contexts involving benefits to others (Lockwood et al., 2016) and heightened sensitivity to others’ pain (Singer et al., 2004). Although psychopathy and empathy are often viewed as opposing socioaffective traits relevant to prosocial motivation and charitable giving (Gunschera et al., 2022; Tusche et al., 2016), little is known about how they shape individuals’ susceptibility to social influence in donation behaviors.

In the current study, we developed a social influence task on charitable giving and implemented it in four pre-registered independent experiments (n1 = 356, n2 = 372, n3 = 375, n4 = 388, total N = 1491). This paradigm enabled us to examine how the statistical properties of others’ donations - specifically their mean and variability - shape individual donation behaviors both in terms of donation magnitude and variability. Furthermore, we applied RL-based modeling to quantify the social information learning process underlying donation behaviors. Finally, we investigated whether individual traits, namely psychopathic traits and empathy, modulate susceptibility to social information. Across all four experiments, we found that observing generous donations led to significant increases in individual giving, whereas observing stingy donations prompted reductions in donation amounts, but that the variance of observed donations did not affect this mean shift. It did, however, affect the variability. Following exposure to others’ donations, the variability of individual donations consistently decreased across conditions. Notably, this reduction in variability was stronger when others’ donations were more consistent than when they were more variable. The effects of both the mean and variability of others’ donations generalized to novel charitable donations. Importantly, psychopathic traits were positively associated with susceptibility to others’ donations, a latent variable derived from computational modeling. This positive association between psychopathy traits and social information use appeared consistently across all studies, in both model-based and model-agnostic behavior, and even generalized to a perceptual decision-making task, where participants had to guess animals on the screen rather than make donation decisions (Molleman et al., 2019). Together, these findings demonstrate that the magnitude and variability of others’ giving shaped the level and variability of individual donations, respectively, and highlight psychopathy as a key factor underlying individual differences in social information use.

Results

Participants completed a novel social influence task on charitable giving (Fig. 1). They were randomly assigned to one of four group-norm conditions: Low Mean-Low SD (LM-LSD) of observed others’ donations; Low Mean-High SD (LM-HSD); High Mean-Low SD (HM-LSD); High Mean-High SD (HM-HSD). Across all four pre-registered experiments, the task consisted of two phases. In the baseline donation phase, participants chose how much they were willing to donate to each of twenty charities, providing a measure of baseline donation tendencies. Participants then entered the observation of others’ donations and second donation phase. This phase comprised twenty rounds involving the same charities. In each round, participants predicted and then observed the donations of five different other donors. Upon completing these five prediction-observation trials, they made a second donation. Donation patterns of others varied across conditions in both generosity (Low vs. High Mean) and consistency (Low vs. High SD), allowing us to quantify how these features of social information influenced individual donation shifts. Measures of psychopathy and empathy were administered after the task to probe their associations with susceptibility to social information.

Figure 1.

In Experiment 1, college students completed a hypothetical donation task with amounts ranging from $0 and $2 (Table 1). To examine the robustness of social influence effects across monetary scales, the donation range was increased to $0-$10 and $0-$100 in Experiment 2 and 3, respectively, with Experiment 3 additionally introducing a larger contrast between variance conditions. Given potential discrepancies between hypothetical and real giving, Experiment 4 employed an incentive-compatible design with a more diverse Prolific sample. This experiment further examined whether social information effects generalized to donations to novel charities in the absence of social information and included a perceptual social influence task to assess the domain generality of associations between personality traits and susceptibility to social influence.

Table 1.

Observing Others’ Generous (Stingy) Donations Increased (Decreased) Individuals’ Donation Amounts, Regardless of the Variance in Social Information

First, we examined whether individual donations were affected by the mean and standard deviation of observed others’ donations. A linear mixed-effects model on individual donation amounts revealed a significant interaction between the mean of others’ donations and the phase across all four experiments (Exp 1: b = -0.34, SE = 0.023, t(12652) = -14.83, p < 0.001, 95% CI = [-0.39, -0.30]; Exp 2: b = -1.83, SE = 0.10, t(13432) = -18.16, p < 0.001, 95% CI = [-2.04, -1.64]; Exp 3: b = -18.03, SE = 1.02, t(13627) = -17.71, p < 0.001, 95% CI = [-20.03, -16.04]; Exp 4: b = -15.42, SE = 0.87, t(14563) = -17.63, p < 0.001, 95% CI = [-17.13, -13.70]; Fig. 2; Table S1). More specifically, pairwise comparison revealed that participants significantly increased their donations after observing generous donations from others, whereas participants decreased their donations after observing stingy donations (Contrast of Baseline and Shift: Exp 1: Under LowMean, ∆ = 0.20, SE = 0.01, z = 17.96, p < 0.001; Under HighMean, △ = -0.16, SE = 0.01, z = -13.35, p < 0.001; Exp 2: Under LowMean, ∆ = 1.03, SE = 0.05, z = 20.34, p < 0.001; Under HighMean, ∆ = -0.91, SE = 0.05, z = -17.88, p < 0.001; Exp 3: Under LowMean, ∆ = 7.75, SE = 0.52, z = 15.06, p < 0.001; Under HighMean, ∆ = -10.80, SE = 0.51, z = -21.22, p < 0.001; Exp 4: Under LowMean, ∆ = 8.15, SE = 0.44, z = 18.50, p < 0.001; Under HighMean, ∆ = -7.42, SE = 0.43, z = -17.20, p < 0.001). The preregistered mixed ANOVA on the averaged individual amount yielded convergent evidence for the same results (Table S2), and also confirmed what we had hypothesized in the preregistrations relating to the effect of the mean. Overall, this pattern demonstrated that participants adjusted their giving behaviors toward the generosity level displayed by others. As a control analysis, we focused on the baseline phase and found no significant differences between conditions (Table S3), suggesting that the subsequent donation changes were driven by exposure to others’ donations rather than by baseline variability.

Figure 2.

In contrast, evidence for an interaction between the standard deviation of observed others’ donation and phase was generally absent, appearing solely (and weakly) in Experiment 3 (Exp 3: b = 2.09, SE = 1.02, t (13627) = 2.05, p = 0.04, 95% CI = [0.09, 4.09]) but not in the remaining three experiments (Exp 1: b = 0.008, SE = 0.02, t (12652) = 0.34, p = 0.7, 95% CI = [-0.04, 0.05]; Exp 2: b = 0.05, SE =0.10, t (13432) = 0.50, p = 0.62, 95% CI = [-0.15, 0.25]; Exp 4: b = 0.33, SE = 0.86, t (14563) = 0.38, p = 0.70, 95% CI = [-1.36, 2.02]). Furthermore, contrary to what we had initially hypothesized, results from the preregistered mixed-effects ANOVA do not show any significant standard deviation and phase interaction in any of the 4 experiments (Table S2). No significant three-way interactions between the mean, standard deviation of others’ donations and phase were present across all experiments (Exp 1: b = -0.03, SE = 0.03, t(12652) = -1.07, p = 0.29, 95% CI = [-0.10, 0.03]; Exp 2: b = -0.20, SE = 0.14, t(13432) = -1.38, p = 0.17, 95% CI = [-0.47, 0.08]; Exp 3: b = -1.04, SE = 1.45, t(13627) = -0.72, p = 0.47, 95% CI = [-3.88, 1.79]; Exp 4: b = -0.30, SE = 1.23, t (14563) = -0.25, p = 0.81, 95% CI = [-2.72, 2.11]). Taken together, these findings indicate that individual donation amounts were robustly shaped by the average generosity of others, with little evidence for the influence of variability of others’ donations.

Effect of Variance on Variance: The Variability of Individual Donations Decreased After Observing Others’ Donations, Especially When Others Showed Less Variability in Their Giving

We then investigated whether individual donation amounts became more or less variable following exposure to others’ giving, and how the average magnitude and variability of others’ donations modulated these changes. Donation variability was quantified as the standard deviation of individual donations within each phase. The results of a mixed ANOVA on donation standard deviation across charities revealed a significant main effect of phase for all experiments (Exp1: F(1, 321) = 146.63, p < 0.001, partial η² = 0.31; Exp2: F(1, 341) = 260.99, p < 0.001, partial η = 0.43; Exp3: F(1, 346) = 144.57, p < 0.001, partial η² = 0.30; Exp4: F(1, 370) = 73.18, p < 0.001, partial η² = 0.17; Fig. 3; Table S4), demonstrating that the donation variability decreased significantly after participants observed others’ donations, and this effect held under all conditions, irrespective of the distribution characteristics of others’ donations. Moreover, a significant interaction between the SD of others’ donations and phase was observed in Experiments 2-4 (Exp2: F(1, 341) = 14.64, p < 0.001, partial η² = 0.04; Exp3: F(1, 346) = 18.44, p < 0.001, partial η² = 0.05; Exp4: F(1, 370) = 14.55, p < 0.001, partial η² = 0.04), while Experiment 1 showed a similar but non-significant pattern (Exp1: F (1, 321) = 2.87, p = 0.09, partial η² = 0.009). Mirroring these results, a two-way ANOVA on the change in donation standard deviation (Shift-Baseline) revealed a significant main effect of others’ donation variability. Post hoc analyses further showed that the variability in individual giving declined to a greater extent after exposure to more consistent (LowSD) than more variable (HighSD) social information in Experiment 2-4 (Contrast of HighSD and LowSD: Exp2: Under LowMean, ∆ = 0.21, SE = 0.10, t (341) = 2.04, p = 0.04; Under HighMean, ∆ = 0.34, SE = 0.10, t(341) = 3.34, p < 0.001; Exp3: Under LowMean, ∆ = 2.46, SE = 1.09, t(346) = 2.25, p = 0.03; Under HighMean, ∆ = 4.15, SE = 1.08, t(346) = 3.83, p < 0.001; Exp4: Under LowMean, ∆ = 2.65, SE = 1.01, t (370) = 2.62, p = 0.009; Under HighMean, ∆ = 2.75, SE = 0.99, t(370) = 2.78, p = 0.006), with Experiment 1 showing a nonsignificant trend in the same direction (Exp1: Under LowMean, ∆ = 0.03, SE = 0.02, t (321) = 1.32, p = 0.19; Under HighMean, ∆ = 0.02, SE = 0.02, t(321) = 1.08, p = 0.28). Thus, while the variance of social information did not seem to affect the mean donation amounts, it did influence the variance of donation amounts, such that individual donations became more consistent after exposure to more consistent others’ donations than to more variable others’ donations.

Figure 3.

In Experiment 1-3, the mean and standard deviation of social information were manipulated at the aggregate level (Fig. S2). Consequently, the variability of within-charity mean donation amounts were greater in the HighSD than in the LowSD conditions, raising the possibility that the observed variability effect may partially reflect a carryover of mean-related influences. To address this concern, Experiment 4 manipulated social information at the charity-item level (See Method S3 and Fig. S2). In addition, we introduced an alternative behavioral index, termed ‘pseudo-SD of individual donation’ (See Method S4) to quantify donation variation relative to social information rather than overall donation dispersion. This alternative analysis produced results consistent with those reported above (Result S1, Fig. S3 and Table S5). Taken together, these findings indicate that the variability cues embedded in others’ donations shaped the distribution of individual givings: observing more consistent donation patterns from others narrowed the variation of individual donations, whereas the exposure to greater variability of others’ donations enabled greater heterogeneity of individual givings.

Generalization of Mean and Variability Effects of Others’ Donations to Novel Charitable Giving

We next assessed whether the effect of the mean and variability of observed others’ donations persisted beyond the immediate prediction-observation phase and generalized to novel charities, even in the absence of social information. In Experiment 4, after completing the observation of others’ donation and second donation phase, participants made five additional donations to novel charities without receiving any further social input.

First, we focused on transfer of the mean effect. We observed a significant main effect of the mean of others’ donation on the donation amount toward novel charities: people in High-Mean conditions donated more than those in Low-Mean conditions (b = 16.94, SE = 3.37, t(370) = 5.03, p < 0.001, 95% CI = [10.34, 23.54]; Table S6). All other effects were nonsignificant, including the main effect of others’ donation variance and the mean and variance interaction.

Second, we investigated whether variability in observed donations modulated the variability of participants’ novel giving, operationalized as the standard deviation of donations to novel charities. This analysis revealed a significant main effect of the others’ donation variability (F(1, 370) = 9.87, p = 0.002, partial n² = 0.026; Table S7). Compared with participants exposed to High-SD social information, those exposed to Low-SD social information exhibited less variability in their novel donations, with this effect being especially pronounced in the Low-Mean condition. No other effects were significant, including the main effect of others’ donation mean and the mean and variance interaction. Overall, these results indicate that the statistical properties of observed donations - the mean and variability - continued to shape, respectively, the magnitude and variability of individual donations in novel contexts.

Figure 4.

The Hybrid Model of Initial Donations and Predictions of Others’ Donations Explains Donation Behaviors Best

To formally examine trial-by-trial learning about others’ donations and subsequent donation changes, we fit participants’ choice behavior to a series of computational models. Specifically, the candidate model set comprised six models (See Methods for full model descriptions):

Model 0 (a non-learning model): This model serves as a baseline, assuming that both participants’ predictions of others’ donations and their second donations are fixed constants.

Model 1 (a D1 Only model): This model assumes that both predictions of others’ donations and individual second donations are solely governed by the individual initial donations.

Model 2 (a Prediction-Only model): This model assumes that participants update their trial-by-trial predictions of others’ donations using a Rescorla-Wagner rule. It also accounts for self-referential bias in the first prediction for each new charity item through a mixture of one’s initial donation and the carryover prediction from the previous item. Second donations are modeled as a linear function of the updated prediction of others’ donations.

Model 2b (variant of the Prediction-Only model): This model is similar to Model 2, except that it removes the self-referential anchoring assumption for the first prediction in each charity item.

Instead, the initial prediction for each new item is set to the final updated prediction from the previous item, yielding a continuous prediction sequence across items. The first prediction of the first charity item is treated as a free parameter. As in Model 2, second donations depend linearly on the updated prediction of others’ donations.

Model 3 (a Hybrid model, combining the D1 and Prediction model): This model adopts the same prediction-learning process as Model 2. Second donations are modeled as a weighted combination of participants’ initial donations and the updated prediction of others’ donations.

Model 3b (variant of the Hybrid model, combining the D1 and Prediction model): Predictions of others’ donations are updated as in Model 2b. As in Model 3a, second donations reflect a weighted average of the initial donation and the updated prediction of others’ donations.

Model fittings were then compared by using the summed AIC (Akaike Information Criterion), the summed BIC (Bayesian Information Criterion) and the PXP (protected exceedance probability). Lower AIC and BIC indicate better fit after penalizing model complexity, whereas higher PXP reflects model dominance corrected for chance-level differences. Model comparison revealed that participants’ behavioral data were best characterized by the Hybrid of D1 and Prediction model (Model 3) in Experiment 1 to 3, which showed the lowest AIC and BIC and highest PXP among all models (Fig. 5 and Table S8). In Experiment 4, though the variant of Hybrid of D1 and Prediction model (Model 3b) exhibited the lowest AIC and BIC, the Hybrid of D1 and Prediction model (Model 3) had a higher protected exceedance probability (PXP = 67.3%) compared to Model 3b (PXP = 32.7%), suggesting stronger evidence for Model 3 at the group level after accounting for chance. Together, these results suggest that individuals used feedback of actual donations of observed others to update their predictions of others’ donation amounts, with the reliance of their own initial donation in predicting each new item diminishing over time. Importantly, the best-fitting model suggests that participants employed a hybrid strategy, combining their initial self-donation tendencies with their predictions of observed others’ giving to guide their second individual donations.

Figure 5.

We conducted model recovery and parameter recovery analyses to evaluate the model identifiability and robustness. In model recovery analyses, across all four experiments, confusion matrices were close to identity matrices, indicating that data generated by a given model was best explained by that same model (Fig. S4; See Method S5 for model and parameter recovery details). Additionally, to assess the robustness of the winning model, we performed parameter recovery analyses. The results showed that recovered parameters from the winning model were reliably associated with the true parameter values across all four experiments (learning rate: r > 0.98; relative weight on the initial donation: r > 0.89; social information weight: r > 0.98; Fig. S5).

To assess whether the winning model could adequately capture the key behavioral signatures present in the real data, we performed posterior predictive check analyses. Simulated data generated using the best-fitting parameters of each participant revealed that the winning model successfully reproduced both the mean-dependent and variance-dependent behavioral patterns. Specifically, the mean of social information modulated shifts in individual donation amounts, while the variability of social information influenced changes in the dispersion of individual donations (See Method S6 and Results S2, including Figs. S6-8).

Psychopathy Was Positively Associated with Susceptibility to Others’ Donations

We subsequently assessed how individual differences in psychopathy modulated the extent to which individuals changed their donations in the light of social information. We examined it by using both model-agnostic and model-based indexes.

We first focused on absolute donation changes, capturing the magnitude of changes in individual donations after observing others’ choices. The analysis revealed that absolute donation changes from the first to the second individual donations were significantly positively correlated with psychopathy scores in Experiment 2 to Experiment 4, with a marginally positive association observed in Experiment 1 (Exp1: Spearman ρ = 0.10, p = 0.071, 95% CI: [-0.01, 0.21]; Exp2: Spearman ρ = 0.17, p = 0.002, 95% CI: [0.06, 0.27]; Exp3: Spearman p = 0.14, p = 0.009, 95% CI: [0.03, 0.24]; Exp4: ρ = 0.19, p < 0.001, 95% CI: [0.09, 0.29]; Fig. S9). Secondly, we focused on the social information weight parameter (w) obtained from the winning model, Model 3 (Hybrid of D1 and Prediction Model). This parameter captured how strongly individuals incorporated predicted donations of others into their second individual donations relative to their initial donations, where higher value denoted stronger reliance on others’ predicted donations. Significant positive associations between such social information weight parameter and psychopathy score were found across all four experiments (Exp1: Spearman ρ = 0.11, p = 0.046, 95% CI: [0, 0.22]; Exp2: Spearman ρ = 0.18, p = 0.001, 95% CI: [0.07, 0.28]; Exp3: Spearman ρ = 0.14, p = 0.011, 95% CI: [0.03, 0.24]; Exp4: ρ = 0.28, p < 0.001, 95% CI: [0.19, 0.38]; Fig. 6).

Figure 6.

These results revealed a robust positive association between the susceptibility to others’ donations and psychopathy. However, it remained unclear whether this relationship was independent of other factors. To address this, we conducted multiple linear regression analyses controlling for potential confounds. We first examined absolute donation changes while controlling for age, gender, and experimental conditions. The positive relationships between absolute donation changes and psychopathy remained significant across all experiments (Exp1: b = 0.002, SE = 0.001, t (318) = 2.00, p = 0.045, 95% CI = [0.000, 0.003]; Exp2: b = 0.011, SE = 0.005, t (338) = 2.28, p = 0.023, 95% CI = [0.001, 0.020]; Exp3: b = 0.140, SE = 0.045, t (343) = 3.06, p = 0.002, 95% CI = [0.049, 0.225]; Exp4: b = 0.140, SE = 0.040, t(367) = 3.56, p < 0.001, 95% CI = [0.064, 0.222]; Table S9). Second, when examining the social information weight parameter, we additionally controlled other modeling parameters from the winning model along with age, gender, and experimental conditions. The social information weight remained a significant predictor of psychopathy scores in Experiment 2 to Experiment 4 with a marginal significant association in Experiment 1 (Exp1: b = 0.002, SE = 0.001, t (315) = 1.70, p = 0.090, 95% CI = [0.000, 0.005]; Exp2: b = 0.004, SE = 0.001, t (336) = 3.04, p = 0.003, 95% CI = [0.002, 0.007]; Exp3: b = 0.004, SE = 0.002, t(341) = 2.43, p = 0.015, 95% CI = [0.001, 0.007]; Exp4: b = 0.007, SE = 0.001, t(365) = 5.27, p < 0.001, 95% CI = [0.004, 0.010]; Table S10). Collectively, these findings revealed a consistent positive relationship between the susceptibility to others’ donation and psychopathy, even after controlling for confounding factors. This pattern indicates that individuals with higher psychopathy scores were more susceptible to observed others’ donations, independent of other covariates.

We also evaluated the link between the social susceptibility to others’ donation and empathy. We found that there were no significant relationships between empathy and absolute donation changes across all experiments (all p > 0.08; Details in Results S3; Fig. S10; Table S11). Similarly, we didn’t find significant associations between empathy and the social information weight parameter in all experiments (all p > 0.06; Fig. S11; Table S12). As preregistered, we also tested the relationship between psychopathy, empathy, and baseline donation levels. The baseline donation amounts were negatively correlated with psychopathy in Experiment 1-3, but this relationship was not significant (and in the opposite direction) in Experiment 4. In contrast, baseline donations were positively correlated with empathy in Experiment 1, 3, and 4, but not significant in Experiment 2 (Details in Results S4; Fig. S12 and Fig. S13).

The Positive Association Between Psychopathy and Social Information Weight Generalized to Perceptual Decision-Making

A robust positive relationship between psychopathy and susceptibility to others’ donations was observed in our charitable giving context. We then examine whether this association reflects a domain-specific effect or a more general tendency to incorporate social information by testing its generalization to a perceptual decision-making task. This test helps determine whether individual differences in social information use are confined to prosocial behavior or instead may reflect broader traits governing how individuals integrate social information when forming judgments. Participants in Experiment 4 additionally completed a perceptual social influence task, the BEAST (Berlin Estimate AdjuStment Task) (Molleman et al., 2019). In this task, participants made an initial estimate of the number of animals shown on the screen, then had the opportunity to revise their estimate after viewing another person’s estimate. The proportional movement toward the other person’s estimate was computed as the social information weight (See Method S7 for more details). We found a significant positive association between psychopathy and the social information weight in this perception social influence task (Spearman ρ = 0.15, p = 0.004, 95% CI: [0.04, 0.25]; Fig. 7), while psychopathy was not related to participants’ initial estimates of the number of presented animals (Spearman ρ = -0.03, p = 0.51, 95% CI: [-0.14, 0.07]; Fig. S14). Furthermore, this positive correlation between psychopathy scores and the social information weight remained significant after accounting for gender and age (b = 0.003, SE = 0.001, t(365) = 3.10, p = 0.002, 95% CI = [0.001, 0.005]; Table S13). These findings indicate a positive relationship between psychopathy and the social information weight that generalized across both social and perceptual domains.

Figure 7.

Discussion

We conducted four pre-registered experiments to investigate how the magnitude and variability of others’ donations influenced individual charitable giving, as well as how initial donations and the susceptibility to social information related to empathy and psychopathy. Across experiments, individuals significantly increased their personal donations after observing generous donations from others and decreased them after being exposed to stingy donations, showing that the amount others offered systematically shaped participants’ own donation magnitudes. However, the variance of others’ donations did not affect this mean shift. Furthermore, participants exhibited a marked reduction in donation variability after being exposed to others’ donations. Individual donations converged more closely on the mean of others’ donations when participants viewed more consistent, as opposed to more variable, others’ donation patterns, indicating that the variability of others’ donations shaped the distribution of individual giving. This social information learning process underlying donation behavior was well captured by an RL-based model, which models participants’ learning about others’ donations through trial-and-error updating, and then determines participants’ second donations by using a weighted average of participants’ initial donations and the learned social estimates. Experiment 4 further revealed that these effects of the magnitude and variation of others’ donations generalized to novel charitable giving, even when no social information was provided. Furthermore, we found that the susceptibility to observed others’ donations was consistently positively associated with psychopathic traits. This association also emerged in a perceptual-based social influence task, indicating a domain-general pattern across decision domains.

Across four experiments, we showed that individual donation amounts significantly increased after viewing generous donations from other donors, and decreased their donations after observing stingy others’ donations. This effect remained stable regardless of changes in the maximum donation range, differences in the variability of others’ donations between consistent and variable group norms, or whether the donation context was hypothetical or incentivized. Previous studies have demonstrated that others’ charitable giving played an important role in shaping individual giving: individuals tended to shift their donation amounts toward the levels donated by others in both laboratory settings and field experiments (Agerström et al., 2016; Chierchia et al., 2020; Feine et al., 2023; Frey & Meier, 2004; Nook et al., 2016; Shang & Croson, 2009). Yet, these studies largely relied on a single piece of social information - either a donation from one other donor or a summary statistics representing a group of previous donors to assess how individuals respond to this single cue. Our study advances previous work by moving beyond a single social information cue to a more dynamic scenario in which individuals were sequentially exposed to diverse donations from multiple donors. Notably, even in this richer information environment, we consistently observed the same directional effects of social information, with individuals calibrating their donations toward the generosity levels of the observed donor group. In addition, evidence that the variance of others’ donations influenced individual generosity levels was minimal, appearing solely (and weakly) in Experiment 3. Opportunistic conformity, characterized by the selective tendency to align with group norms when it serves individual interests, has been documented in several studies (Charness et al., 2019; Chierchia et al., 2020; Dimant, 2019). In our design, such a pattern would be reflected by larger decreases after observing stingy donations in the variable (vs. consistent) condition, and attenuated increases in donations after observing generous donations in the variable (vs. consistent) condition. However, we found limited evidence for opportunistic conformity in our study. Instead, participants tended to adjust their donations toward the donation levels of others regardless of whether observed donations were consistent or variable. Our findings thus indicate that the magnitude of observed donations might exert a more robust influence on the level of individual donations. Whether the observed donations were generous or stingy, consistent or variable, or presented in hypothetical or incentivized contexts, people reliably shifted their donations toward the observed generosity level, underscoring the magnitude of observed donations as the key determinants of changes in individual donation levels.

We observed a significant decrease in the variability of individual donations following exposure to others’ donations. Notably, this effect was more pronounced when observing more consistently, compared to more variable donation patterns. One possible explanation is that this pattern may arise as a by-product of magnitude effects, such that participants shifted their donations toward the observed levels in both Low- and High-SD conditions, naturally producing corresponding differences in donation variability. We ruled out this possibility through both experimental controls and analytic approaches. In the final experiment, the mean of five others’ observed donations for each item was held constant across Low- and High-SD conditions, eliminating magnitude differences as a confound. The phase-by-variability interaction persisted, despite this constraint. Additionally, we employed a novel behavioral measure, the pseudo-SD of individual donation, reflecting how closely individual donations cluster around others’ mean donations. The Low-SD conditions yielded a greater reduction in pseudo-SD, demonstrating stronger alignment under consistent information. The degree to which others’ opinions are homogeneous or heterogeneous influences our attitude toward their judgments (Pfänder et al., 2025), and potentially impacts personal judgments and behaviors (Brown et al., 2022; Ecker et al., 2022; Li et al., 2025; Molleman et al., 2020). Our results suggest that observing more heterogeneous charitable donations led to greater variability in individuals’ own donation amounts. These findings align with accumulating evidence that low group variability was perceived as a strong and clear social signal, thereby reducing variability in individual behaviors (Brown et al., 2022; Dimant et al., 2024; Li et al., 2025). Prior work has shown that homogeneous group norms exert strong conformer pressure, pulling individuals toward the normative center, whereas heterogeneous group norms provide weaker cues and lead to more dispersed response, such as in the case of cooperation (Dimant et al., 2024). Our results extended this pattern to charitable giving and to contexts in which group norms must be learned incrementally rather than presented all at once. In our design, the social information follows normal distributions, but an important next step will be to examine whether similar effects occur in situations where observed donations follow other distribution forms. Future experimental designs that incorporate skewed or asymmetric distribution might be essential for determining how the shape and distribution of social information impact individual behaviors.

Higher levels of psychopathy have been linked to heightened antisocial behaviors and attenuated prosocial behaviors (Blair et al., 2006; Marsh, 2019; Rhoads et al., 2025). In line with these prior findings, our initial donation results revealed an inconsistent but mainly negative correlation between psychopathy and the initial amount individuals donated to charities in the absence of social cues, whereas psychopathy was unrelated to participants’ initial estimates of the number of animals. Overall, these findings converge with prior evidence suggesting that individuals in higher psychopathic traits tend to be less intrinsically altruistic (Driessen et al., 2021; Gong et al., 2019; Gunschera et al., 2022; Sakai et al., 2019). Given the lower prosociality, we had initially expected that participants higher in psychopathic traits would be less susceptible to social information. Instead, we found that self-reported psychopathic traits were positively associated with the extent to which individuals incorporated social information into their donation decisions. Notably, this positive relationship remained stable across hypothetical and incentivized settings and was observed in both college student and general population samples. Previous research has suggested that people with high psychopathy were more susceptible to peers’ opinions to behaviors, from evaluations of facial attractiveness, institutional misconduct to drug uses (Curtis et al., 2020; Overgaauw et al., 2019; Tatar II et al., 2016). Individuals with elevated psychopathy were grandiose, interpersonally manipulative and superficially charming. People exhibiting psychopathic traits may strategically display adaptive behaviors in response to cues during interpersonal interactions, often to maximize self-interests or achieve desired social goals (Brazil et al., 2025; Gervais et al., 2013; Osumi & Ohira, 2017). Consistent with these findings, our results might suggest that individuals with higher psychopathic traits strategically aligned with observed others’ donations in order to blend in socially, thereby avoiding standing out, facilitating manipulations and strengthening their sense of fitting into the group (Brazil et al., 2023; Glenn et al., 2017; Reale et al., 2020). By adjusting individual donations toward those observed from others, they may sustain a favorable social image while still pursuing their own self-interests in the absence of explicit social cues. Such superficial compliance with social norms is consistent with how individuals high in psychopathic traits display socially desirable actions solely as a strategic tactic rather than from genuine motivations (Gunschera et al., 2022; Lin et al., 2025; Osumi & Ohira, 2017). Future studies could examine the neural mechanisms underlying social influence in individuals with elevated psychopathic traits, as identifying the key neural circuits involved would provide a more mechanistic account of why these individuals respond differently to social information. Additionally, despite the robust and replicable positive association observed between psychopathy and susceptibility to others’ donations, the current study was conducted with non-clinical samples, primarily college students and individuals from the general population. Future research could investigate whether these effects extend to clinical or forensic populations, where the psychopathic traits are likely more severe.

We found that the positive link between psychopathic traits and the degree to which individuals use social cues in their decisions generalized to the perceptual domain. Our results suggest that the positive association between psychopathic traits and the susceptibility to social information emerged not only in the domain of charitable giving where individuals could infer the descriptive norm, but also in the perceptual decision-making task, where individuals merely observed others’ choices. Together, these findings suggest that psychopathy may be associated with heightened domain-general susceptibility to social information - a pattern observed in both norm-driven and ambiguity-driven settings. Nevertheless, it is worth noting that the perceptual task was only done in one of the studies (study 4), and thus the perceptual task psychopathy results necessitate replication. In addition, the two paradigms employed in this study differed substantially. Future research could develop directly comparable designs across domains to better investigate the domain-specific vs. domain-general cognitive and computational mechanisms underlying social influence and how these relate to psychopathy (Lee & Chung, 2022; Olsson et al., 2020; Ruff & Fehr, 2014). Furthermore, both tasks involved two individual decisions - one made before and one after exposure to social information. This structure may have introduced demand characteristics, leading participants to adjust their response in line with perceived experimental expectations (Corneille & Lush, 2023; Iarygina et al., 2025; Orne, 1962). Individuals high in psychopathy may engage in strategic impression management, shifting their behaviors to align with these perceived expectations as a means of influencing how they were viewed by the experimenter (Gillard & Rogers, 2015; Hart et al., 2019). Future studies could explore whether the link between psychopathy and social susceptibility persists across tasks that vary in the salience of demand characteristics, from overt social cues to more subtly framed designs with immersive cover stories.

Together, we find evidence that the statistical properties of observed others’ donations shaped corresponding features of individual donations: individuals increased (or decreased) their donation levels after observing generous (or stingy) levels of others’ donations, while consistent patterns of observed donations yielded a stronger reduction in individual variability than variable patterns. Computational modeling successfully captured the key behavioral patterns underlying changes in charitable giving. Moreover, the susceptibility to observed donations was positively associated with psychopathic traits. These findings remained robust regardless of donation ranges, whether the discrepancy between consistent and variable observed donations was changed, whether giving was hypothetical or incentivized, or whether participants were from student or general population samples. Additionally, a positive association between social information use and psychopathic traits also appeared in a perceptual-based social influence task. These findings could have important implications for our understanding of how specific statistical features of social information systematically shape individual giving behaviors, and how individual differences, particularly in psychopathic traits, may drive variability in this process.

Our results on the mean and variability of individual donations carry direct implications for promoting charitable giving: displaying generous and consistent donor behavior may represent a simple but effective strategy for nudging people toward greater prosocial engagement. Our findings of a consistent positive relationship between psychopathic traits and susceptibility to social information also point to a practically meaningful insight: despite the lower intrinsic prosociality in individuals with elevated levels of psychopathy, presenting generous donations from others may serve as an actionable pathway for promoting prosocial behavior in this population. Overall, our results may thus directly inform the development of more effective interventions to promote charity giving at both the individual and group levels, with potential insights to promote prosocial behavior more generally.

Methods

Participants

Participants in Experiment 1-3 were recruited online through the course extra credit system at the University of Minnesota. Participants had to be at least 18 years old and have normal or corrected-to-normal vision. To obtain a more diverse general population sample, participants in Experiment 4 were recruited through Prolific Academic (https://www.prolific.com). Participation criteria included (1) United States as the country of residence, (2) age between 18 and 80 years, (3) fluency in English, and (4) normal or corrected-to-normal vision. The sample size was pre-registered, and determined based on prior work (Nook et al., 2016), which reported an effect size for the main effect of the mean of others’ donations (f = 0.23). To account for potential publication biases, we performed an a priori power analysis in G*Power 3.1 (Faul et al., 2009) using a slightly smaller effect size (f = 0.20). This analysis indicated that a minimum of N = 327 participants was required to achieve 95% power at a significance level α = 0.05. All participants across studies completed the experiment online. All experiments were preregistered (Exp1: https://aspredicted.org/9jwt-48nh.pdf; Exp2: https://aspredicted.org/4ys2ws.pdf; Exp3: https://osf.io/cz2u3; Exp4: https://osf.io/fcugq).

We recruited 356 participants in Experiment 1. As preregistered, 31 participants who failed more than one of seven attention checks were excluded (See Method S1). Data of 325 participants were thus analysed (236 females, 81 males, 8 classified themselves as non-binary, other or prefer not to respond; age = 20.24 ± 3.49 [M ± s.d.]). Experiment 2 included 372 participants. 27 participants were excluded based on the same exclusion criterion as Experiment 1, resulting in the data of 345 participants being analysed (285 females, 55 males, 5 classified themselves as non-binary, other or prefer not to respond; age = 20.05 ± 2.77 [M ± s.d.]). Experiment 3 had 375 participants, with 25 participants excluded based on the same exclusion criteria, leaving 350 participants for analysis (285 females, 56 males, 9 classified themselves as non-binary, other or prefer not to respond; age = 20.19 ± 2.78 [M ± s.d.]). Experiment 4 included 388 participants, of whom 14 were excluded according to the criteria, so 374 participants remained in the formal analysis (218 females, 151 males, 5 classified themselves as non-binary, other or prefer not to respond; age = 41.16 ± 12.73 [M ± s.d.]).

In Experiment 1-3, participants were awarded extra course credit upon completion of the experiment. In Experiment 4, participants received a fixed payment (about $10/hour), plus a bonus of up to $2 based on their decisions in the donation task and the perceptual animal counting task. One of their donations was randomly selected for actual implementation. Our study protocol has been approved by the University of Minnesota Institutional Review Board and all participants provided informed consent before participating in the study.

Task Design - Donation Task

To investigate how the mean and variability of other’s donations influenced individual givings, we employed a 2 (Mean of the observed donation: Low vs. High) x 2 (Standard Deviation (SD) of the observed donation: Low vs. High) between-subjects design. Participants were randomly assigned to one of these four group-norm conditions (e.g. the High-Mean-Low-SD of the observed donation condition). All participants completed two task phases in sequence: (1) The baseline donation phase; (2) The observation of others’ donation and individual second donation phase. The experimental settings of Experiment 1 are elaborated in detail first, whereas variants across different studies are outlined subsequently. The task itself was programmed using jsPsych (de Leeuw, 2015) and run on the cognition.run platform (https://www.cognition.run/), and all participants performed the task online, in their own devices.

In the baseline donation phase, participants were asked to decide how much they were willing to donate to each of 20 different charities, from $0 to $2 in Experiment 1. These 20 charities were dedicated to missions in education, healthcare, environment and societal issues. The inclusion of diverse charities was used to minimize confounds from idiosyncratic charity features in measuring prosocial-influence effects (Chierchia et al., 2020). In each round, participants were first shown a brief description of a charity alongside its logo, included to enhance task engagement. To discourage uninformed decisions, the donation slider was displayed only following a 3-second presentation of charity information. Participants were then able to indicate their donations by adjusting the slider on a visual analogue scale ranging from $0 (donate nothing) to $2 (donate all). The slider’s starting value was concealed until participants made their first adjustment, to avoid anchoring effects. All decisions were self-paced with no time limit. Each donation was followed by a fixation cross displayed for a randomly selected duration of 0.75, 1, or 1.25 s. The order of charities was randomized across participants. The baseline phase served to quantify individual donation propensities in the absence of any experimental manipulations.

In the observation of others’ donation and individual second donation phase, participants predicted and observed donations from multiple others, and were then requested to donate again. In each round, participants were first asked to predict the first person’s donation to a particular charity without a time limit. Their prediction and that person’s actual donation were then displayed for at least 2 s. This prediction-then-feedback sequence was repeated for four additional others, after which participants donated to the same charity again. In total, participants completed 20 rounds (one per charity): five prediction-feedback trials per charity (100 predictions overall) followed by one individual donation per charity (20 donations total). The charity set was identical to the baseline donation phase, but was presented in a randomized order. We used jittered fixation: 0.75/ 1.0/ 1.25 s between prediction-feedback trials, and 2.25/ 2.5/ 2.75 s before the donation trial and between rounds. The first and second individual donations were probed in separate phases to allow us to quantify how individual donation changes as a function of the social-information manipulations. All presented donations from others were choices collected from real participants. Social information in Experiment 1 was obtained from a pilot study that collected real donation amounts from participants (see Method S2 in SI). In Experiment 2-4, social information was derived from both the pilot study and prior experiments with donation values rescaled to match the corresponding range of each experiment. For instance, Experiment 2 used social information from Experiment 1 and the pilot data, with others’ donation amount multiplied by five to match the expended range. All participants were fully debriefed after the experiment regarding the source and nature of social information.

In Experiment 1, participants made hypothetical donations within a $0-$2 range (Fig. S1A). To construct the Low-Mean and High-Mean groups, we chose representing means of M = 0.6 (Low-Mean) and M = 1.4 (High-Mean), respectively, and standard deviations of SD = 0.1 (Low-SD) and SD = 0.3 (High-SD). Because we selected donation amounts based on actual participant contributions, this introduced slight differences between the conditions (Fig. S1A; Low-Mean-Low-SD condition: Mean = 0.593, SD = 0.103; Low-Mean-High-SD condition: Mean = 0.604, SD = 0.284; High-Mean-Low-SD condition: Mean = 1.400, SD = 0.115; High-Mean-High-SD condition: Mean = 1.401, SD = 0.292). Experiment 2 expanded the range to $0-$10 and proportionally scaled social information from Experiment 1 to preserve relative means and variances (Fig. S1B; Low-Mean-Low-SD condition: Mean = 3.00, SD = 0.514; Low-Mean-High-SD condition: Mean = 3.02, SD = 1.418; High-Mean-Low-SD condition: Mean = 7.00, SD = 0.576; High-Mean-High-SD condition: Mean = 7.00, SD =1.460). Donation remained hypothetical. In Experiment 3, the donation range was $0-$100, and the variability contrast between Low-SD and High-SD conditions was increased. Donations were still hypothetical (Fig. S1C; Low-Mean-Low-SD condition: Mean = 29.76, SD = 4.035; Low-Mean-High-SD condition: Mean = 30.74, SD = 16.089; High-Mean-Low-SD condition: Mean = 70.30, SD = 4.049; High-Mean-High-SD condition: Mean = 69.42, SD = 16.466). To encourage real and honest responses (Camerer & Mobbs, 2017), Experiment 4 was incentive-compatible. Donation was recorded on a 0-100-points scale (100 points = $1) (Fig. S1D; Low-Mean-Low-SD condition: Mean = 29.65, SD = 4.191; Low-Mean-High-SD condition: Mean = 29.65, SD = 15.456; High-Mean-Low-SD condition: Mean = 69.65, SD = 4.191; High-Mean-High-SD condition: Mean = 69.65, SD = 15.456). For cleaner specification of SD effects, the per-charity mean of others’ donation was matched across Low-SD and High-SD conditions (See Method S3).

Procedure

In Experiment 1-3, participants first provided informed consent and read the instructions for the baseline donation phase. They then were asked to fill out demographic information (including age, gender, race, ethnicity and highest level of education). Before entering the formal task, they performed two practice donation rounds and one practice attention check round. Upon completion of the first phase, they were directed to instructions of the second phase, which involved observing others’ donations and making their second donations. They then completed practice trials (two prediction trials, one individual donation trials and one attention check), before proceeding to the formal task of the second phase. At the end of the behavioral task, participants estimated the range of others’ donations, and then indicated the average donation of all donors they encountered in the second phase, along with their confidence in this estimate. Afterward, they completed two questionnaires: the Levenson’s Self-Report Psychopathy Scale (Levenson et al., 1995) and the Questionnaire of Cognitive and Affective Empathy (QCAE) (Reniers et al., 2011).

In Experiment 4, the procedure was the same as above with the difference that participants completed demographic information right after giving consent. Also, after completing the Levenson’s Self-Report Psychopathy Scale and the QCAE, they completed a brief perceptual-based social influence task (See Method S7) and an additional set of questionnaires not examined here.

Behavioral Data Analysis

To examine how the mean and variability of observed others’ donations influenced individual donation tendencies, we fitted a mixed-effect linear regression, with trial-level donation amounts as the dependent variable. The independent variables included the mean of observed donation (Low vs. high), the SD of observed donation (Low vs. High), the donation phase (baseline vs. observation of others’ donation and individual second donation), as well as all two-way and three-way interactions among these factors. Random intercepts were included for both participants and charity items, to account for repeated measures. The model was as formally specified as:

Donation Amount ~ MeanCond (Low vs. High) * SDCond (Low vs. High) * Phase (Baseline vs. Observation) + (1| Item) + (1| Subject)

We also conducted the pre-registered 2 (Mean of observed donation: Low vs. High) × 2 (SD of observed donation: Low vs. High) × 2 (Phase: Baseline vs. Shift) mixed-design ANOVA on participants’ averaged donation amount.

To assess whether the average level and spread of observed others’ donations modulate the variability of individual donations, we calculated the standard deviation of each participant’s donations (across charities) separately for each phase. These values were then entered into a mixed-design ANOVA, with the Mean of other’s donation (Low vs. High) and SD of others’ donation (Low vs. High) as between-subjects factors, and Phase (Baseline vs. Observation) as a within-subjects factor.

To investigate the association between psychopathy and donation changes after exposure to others’ donations, we computed Spearman correlations between the total score of Levenson’s Self-Report Psychopathy Scale, QCAE and the absolute difference between participants’ initial and subsequent donations.

Computational Modeling

To quantitatively understand how individuals learned about others’ donations and changed their individual donations in response to observed information, we built a series of models and compared their fit to participant’s choice data.

Model 0 (Intercept Only Model)

Model 0 assumes that participants predict other’s donations using a fixed constant, and their second donation did not vary.

Model 1 (D1 Only Model)

Model 1 assumes that participants predict other’s donations and update their own second donations using the linear functions of their initial donations. It means that both predicted others’ donations and second donations are modeled as regressions with the constant intercept and the slope on the initial donation.

Model 2 (Prediction Only Model)

Model 2 employs the Rescorla-Wagner rule to model individuals’ predictions of other’s donations. Specifically, this model uses the prediction errors (PE) and the current estimate of others’ donation to update the prediction of others’ donation amount (Pred) in the next trial. The prediction error (PE) is defined as the difference between the feedback of other’s actual donations and the prediction on that trial.

The prediction of others’ donation is updated based on the following equation:

The learning rate parameter (α) specifies the extent to which the new information is integrated into the current beliefs (ranging between 0 and 1).

To account for the potential self-referential bias in the first prediction of each charity item, this first prediction is modeled as a mixture of the participant’s initial donation and the predicted donation from the last trial of the previous charity item. Furthermore, the relative weight on the initial donation is modeled as λ(^i–1), indicating it decreases geometrically with the number of charity items (i). At the first time (i = 1), the prediction is determined entirely by individual initial donations; as the item number (i) increases, its weight decreases, while the carryover prediction term receives an increasing complementary weight. λ is bounded between 0 and 1. Larger λ indicates that people place more weight on their own initial donation when forming the first prediction for a new item; a smaller λ means predictions are primarily driven by the estimate of others’ donations from the prior item.

In Model 2, it assumes that the second donation depends completely on the prediction of others’ donations. Specifically, the second donation is modeled as a regression on the last prediction of others’ donation within each item, with the intercept and slope term.

Model 2b (Prediction Only Model-Free Initial Prediction)

Model 2b mirrors Model 2, except that it omits the assumption of self-reference bias when predicting others’ donations. It still allows the use of the Rescorla-Wagner rule to update predictions of others’ donations. Instead, it assumes that the prediction process is modeled as a single continuous sequence: the final updated prediction from the previous item (i) becomes the initial prediction for the next item (i + 1). In addition, we assume that each person has an individual initial prediction about the other’s donation. Therefore, the first prediction of others’ donation is set as a free parameter (bounded between 0 and the donation range).

Consistent with Model 2, the second donation is specified as a linear function of the participant’s prediction of others’ donations, with no additional predictors.

Model 3 (Hybrid of D1 and Prediction)

In modeling the prediction of others’ donation, we adopt the same updating rule as Model 2: the prediction of others’ donations is updated via the Rescorla-Wagner rule, and the first within-item prediction incorporates their initial donation geometrically over time.

For the second donation, Model 3 assumes that individuals combine both their initial donations and their prediction of observed others’ donation to guide their second donations. Accordingly, the second donation is modeled as a weighted average of their initial donation and the prediction of other’s donation. This weighted average is governed by a weighting parameter (w) bounded between 0 and 1, specifying the extent to which the second donation is based on their own initial donation (w = 0) versus prediction of observed donations from others (w = 1).

Model 3b (Hybrid of D1 and Prediction-Free Initial Prediction)

Like Model 2b, it assumes that updating predictions of others’ donations follows the Rescorla-Wagner rule and predictions are treated as a continuous across items - the final updated prediction for one item (i) serves as the initial prediction for the next item (i + 1).

As in Model 3, the second donation depends on both participant’s initial donation for the item and the predicted donation of others via a weighted average.

Model fitting and model comparison

All models were fitted to participants’ behavioral responses to identify the model that best characterized participants’ choices. Parameter optimization was conducted by using nonlinear constrained minimization (implemented in the Matlab function fmincon), using 500 random start points for parameters to find the best-fitting parameters for each individual. This procedure was employed to minimize the sum of squared errors (SSE) between model-predicted and actual responses across trials.

To evaluate the model performance, we used two complementary information criteria: the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Both criteria penalize goodness-of-fit by the model complexity. They were calculated as follows:

where n denotes the number of observations, and k represents the number of free parameters. Lower AIC and BIC indicate better model fit after accounting for model complexity. Both AIC and BIC scores were summed across participants for each model.

In addition, the posterior exceedance probability among all models was computed. This measure is used to estimate the probability that any model is more frequently used in the population than other models within the model set, while correcting for the possibility that observed differences were driven by chance rather than true population-level effects. Consistent with previous studies (Doyon et al., 2025; Frolichs et al., 2022), random-effect family-wise model comparison was conducted using the Bayesian Model Selection (BMS) procedure in the SPM12 Matlab toolbox (function spm_BMS) (Rigoux et al., 2014).

Model recovery and parameter recovery analyses were conducted to evaluate the identifiability and reliability of candidate models and estimated parameters. Detailed procedures are provided in the Supplementary materials (Method S5).

Use of AI

During the manuscript preparation, ChatGPT (GPT 5.3, OpenAI) was used, solely for language polishing and wording refinement. Prompts were used for this purpose, including ‘help me revise this sentence in several versions’, and ‘make this sentence more concise’. All outputs were carefully reviewed by the authors. These AI tools were not used to generate statements, design experiments, interpret findings or make scientific conclusions.

Supplementary Materials

Supplementary Methods

Method S1. Attention check questions

As preregistered, seven attention check items were included for each experiment. These consisted of two catch trials in the baseline donation phase, three during the observation of others’ donation and individual second donation phase, and two within the questionnaire section. The specific items are listed below.

In the baseline donation phase, attention checks (using Experiment 1 as an example; numeric ranges were proportionally changed according to the maximum donation range in Experiment 24) were:

1. This is a forced choice round: Please choose a number between $0.6 and $1.

2. This is a forced choice round: Please choose a number between $1.2 and $1.6.

In the observation of others’ donation and individual donation shift phase, the following terms were used:

1. This is a forced choice round: Please choose a number between $1.6 and $2.

2. This is a forced choice round: Please choose a number between $0 and $0.4.

3. This is a forced choice round: Please choose a number between $0.6 and $1.

Within the questionnaires (empathy and psychopathy scales, respectively), two additional attention check questions were included:

1. Please select ‘Strongly Disagree’ for this question.

2. For this item please choose ‘Agree somewhat’.

Method S2. Pilot study

To obtain social information presented to participants in the social influence task on charitable giving, we conducted a pilot study with college students (N = 169; 120 females, 45 males, 4 identifying as non-binary; age = 19.98 ± 2.10 [M ± s.d.]). Participants received extra credits as compensation. In the pilot study, participants indicated their willingness to donate to a series of charities, with donation amounts ranging from $0 to $2. Following the donation task, participants completed the Questionnaire of Cognitive and Affective Empathy (QCAE) (Reniers et al. 2011). Donation amounts from the pilot study were presented to participants in the formal task either in their original values or after being multiplied by a constant factor. In the pilot study, individual donation amounts were significantly positively associated with empathy (b = 0.007, SE = 0.002, t(146) = 2.80, p = 0.0006, 95% CI = [0.002, 0.012]).

Method S3. Item-level social information settings in Experimental 4

In Experiment 1-3, we systematically manipulated the mean and standard deviation of the social information across four conditions. At the aggregate level, mean donation amounts were matched between LowMean-LowSD and LowMean-HighSD conditions, as well as between HighMean- LowSD and HighMean-HighSD conditions. Likewise, overall standard deviations were matched between LowMean-LowSD and HighMean-LowSD conditions, and between LowMean-HighSD and HighMean-HighSD conditions (Fig. S1). However, these constraints were implemented at the overall level rather than at the charity-item level (Fig. S2). Consequently, the observed variability effect (participants showed lower variability in their second donation decisions compared to their first donation decisions, and more so in the lowSD conditions) may partially reflect a carryover of mean-related effects: the dispersion of the within-charity mean donation amounts tended to be greater in HighSD than in LowSD conditions. Under this structure, when participants shifted their donations toward the mean social information for each charity, the standard deviation of their donations may be expected to be larger in HighSD conditions than LowSD conditions.

To rule out this possibility, we manipulated the social information at the charity-item level in Experiment 4. Specifically, within each charity item, we held the average of the five observed donations constant between LowSD and HighSD conditions within both the LowMean and HighMean settings, while ensuring that the standard deviation of the five observed donations was higher in HighSD than in LowSD conditions. This design guarantees that the mean level of others’ donations was identical at the item level across SD conditions, whereas the variability differed systematically, allowing us to test variability effects without confounding from mean differences.

Method S4. Analysis for testing the variability effect

We further evaluated whether the variability of social information influenced changes in participants’ donation variability. Importantly, instead of focusing exclusively on the variability of participants’ donations per se, we assessed the extent to which participants’ donation converged toward (or diverged from) the social information provided for each charity item.

To this end, we derived a behavioral index termed ‘pseudo-SD of individual donation ’. This index is conceptually analogous to the conventional standard deviation, but instead of centering each participant’s donation around their own mean across items, it centers individual donations around the mean donation of five observed donors for the corresponding charity item.

This measure provides the degree of variation in participants’ donations around the mean of others’ donations, with lower pseudo-SD values indicating reduced variation from others’ donation levels and higher values indicating increased variation across items.

We then conducted a mixed-design ANOVA on pseudo-SD of individual donation, mirroring the analysis performed on the SD of individual donation. This analysis allows us to examine the variability effect in a way that is less susceptible to potential mean-induced differences in donation dispersion, while still characterizing the degree of variation in donation behavior.

Method S5. Model and Parameter recovery

To evaluate whether data simulated under a given model would be best fit by the same model, we constructed confusion matrices for all four experiments. For each confusion matrix, parameter values were randomly drawn (uniformly) from predefined ranges to simulate data 360 times (85 artificial agents per condition) under each candidate model. Gaussian noise was added to the simulated data (SD = 0.1, 0.5, 5 and 5 respectively, based on the range of 2, 10, 100, and 100 in the corresponding experiment). All candidate models were subsequently fitted to the simulated data, and model recovery was assessed using BIC. Confusion matrices were based on the probabilities of selecting each candidate model as the best-fitting model for data generated by a particular model, using BIC values, with ideal recovery reflected by dominant diagonal elements.

In addition, we conducted parameter recovery analyses to evaluate whether parameters of the winning model could be reliably recovered. For each experiment, we simulated data of 360 artificial agents (85 agents per condition) using randomly sampled parameters of the winning model. Gaussian noise was added to the simulated data (SD = 0.1, 0.5, 5, and 5, respectively). The winning model was then fitted to the simulated datasets to obtain recovered parameter estimates. Parameter recovery was quantified by computing correlations between the groundtruth and recovered parameters, with higher values reflecting better recovery.

Method S6. Posterior predictive checks

To evaluate whether the winning model (Model 3: a Hybrid model, combining the D1 and Prediction model) could reproduce the behavioral signatures present in the real data, we conducted the predictive check analyses. First, the predicted second donation amounts were generated using each participant’s best-fitting parameters from the winning model. Second, the simulated data were analyzed using the same procedures applied to the empirical data. Linear mixed-effects models were fitted to donation amounts, whereas mixed-effects ANOVAs were conducted on the SD of individual donation amounts and the Pseudo SD of individual donation amounts. First donation amounts were retained from the empirical data.

Method S7. The description of the perceptual-based social influence task

At the end of Experiment 4, participants completed a perceptual-based social influence task: the BEAST (Berlin Estimate AdjuStment Task) (Molleman, Kurvers, and van den Bos 2019). This task comprised five rounds presented in a fixed order. In each round, participants viewed an image containing 50-100 animals. After 6 seconds, the image disappeared, and participants were asked to provide an initial estimate of the number of animals (E1 ) within 15 seconds. After entering their first estimate (E1 ), participants were shown social information (X), which was an estimate provided by another person who had previously completed the task. Participants then made their second estimate (E2 ) within 45 seconds. During the second estimate, both their first estimate and the social information were displayed on the screen.

The distance between E1 and X was controlled by selecting social information from a prerecorded pool of estimates provided by 100 previous participants. Across Round 1-5, the target proportional deviation of X from E1 (Δ)) was set to 25%, 15%, 20%, 15%, 25%, respectively. Importantly, the social information always pointed to the direction of the true value. To implement this manipulation, we computed the ‘target’ value (X′) as follows: If E1 was below the true value, X′ = E1 * (1 +Δ) ; If E1 was above the correct value, X′ = E1 * (1 —Δ) ; The displayed social information was then selected as the estimate from the pre-recorded pool that was closest to the target value (X′).

The social information use (s) in each round was calculated as a relative change toward social information: s = (E₂ — E₁)/(X — E₁). Rearranging this equation yielded:E₂ = (1 — s) * E₁ + s * E₁, indicating that the second estimate (E₂) in each round reflected a weighted average of an individual initial estimate (E₁) and the social information (X). s represented a relative weight on social information.

Following previous studies (Gaule et al. 2022; Molleman et al. 2019), we excluded rounds in which participants updated their estimate in the opposite direction of the social information (s < 0) or moved beyond the social information (s > 1). We also excluded rounds in which the first or second estimate was more than 100 animals away from the true value, which likely reflected data entry errors. Rounds were also removed when either estimate was 0 or when participants’ estimates exactly matched social information. In total, 92.74% of rounds were included in the final analysis.

Supplementary Results

Results S1. Assessing the variability effect on the dispersion of participants’ donations around social information

To further probe the robustness of the variability effect and to rule out the possibility that it merely reflects mean-driven differences across social information settings, we devised an alternative behavioral index capturing the dispersion of individual donations relative to others’ donations, termed the ‘pseudo-SD of individual donation’ (See Method S4 for details). It parallels the conventional standard deviation, with the mean of individual donations substituted by the mean donation of five others to the specific charity. Accordingly, the pseudo-SD provides a quantification of how strongly participants’ donations deviated from the social information of each charity.

A mixed ANOVA on pseudo-SD revealed robust main effects of phase across all experiments Specifically, pseudo-SD significantly decreased from baseline to observation phase, indicating that participants’ donations became less dispersed around others’ donation after exposure to social information (Expl: F(1, 321) = 372.79,p < 0.001, partial η² = 0.54; Exp2: F(1, 341) = 464.22, p < 0.001, partial η = 0.58; Exp3: F(1, 346) = 397.08,p < 0.001, partial η² = 0.53; Exp4: F(1, 370) = 238.73, p < 0.001, partial η² = 0.39; Table S5; Fig. S3). Critically, across Experiments 2-4, the interaction between the variability of others’ donations and phase was significant (Exp2: F(1, 341) = 7.82, p = 0.005, partial η² = 0.02; Exp3: F(1, 346) = 10.43, p = 0.001, partial η² = 0.03; Exp4: F(1, 370) = 10.72, p = 0.001, partial η² = 0.03), with Experiment 1 showing a marginal trend in the same direction (Exp1: F(1, 321) = 3.56, p = 0.06, partial η² = 0.01). Specifically, individual donations clustered more closely around the mean of five others’ donations after exposure to more consistent (LowSD) vs. more variable (HighSD) donations from others. Thus, the observed variability effect appears robust to alternative analyses and design constraints, supporting a robust role of social information variability in affecting the variability of individual giving.

Results S2. Posterior predictive predictions

To further determine the extent to which the winning model (Model 3: a Hybrid model, combining the D1 and Prediction model) could account for the behavioral patterns observed in the empirical data, we performed the posterior predictive check analyses. We first examined the predicted donation amount, and the linear mixed-effects model applied to the simulated data showed a significant interaction between the mean of others’ donations and the phase across all four experiments (Exp 1: b = -0.29, SE = 0.021, t(12613) = -13.96,p < 0.001, 95% CI = [-0.34, - 0.25]; Exp 2: b = -1.68, SE = 0.09, t(13432) = -18.06,p < 0.001, 95% CI = [-1.87, -1.50]; Exp 3: b = -15.66, SE = 0.94, t(13627) = -16.74,p < 0.001, 95% CI = [-17.49, -13.82]; Exp 4: b = - 14.44, SE = 0.79, t(14563) = -18.27,p < 0.001, 95% CI = [-16.00, -12.90]; Fig. S6). These results indicate that the effect of the average level of social information on individual donation amounts was replicated in the simulated data. We next focused on the variability effect. Mixed-effect ANOVAs on the SD of predicted individual donation amounts revealed a significant main effect of the phase across all four experiments (Expl: F(1, 320) = 833.875,p < 0.001, partial η² = 0.72; Exp2: F(1, 341) = 927.77,p < 0.001, partial η² = 0.73; Exp3: F(1, 346) = 766.36,p < 0.001, partial η² = 0.69; Exp4: F(1, 370) = 443.04,p < 0.001, partial η² = 0.55; Fig. S7). The interaction between the phase and the variability of others’ donations were also reproduced in Experiments 2-4 (Exp2: F(1, 341) = 14.64,p < 0.001, partial η² = 0.04; Exp3: F(1, 346) = 18.44, p < 0.001, partial η² = 0.05; Exp4: F(1, 370) = 14.55, p < 0.001, partial η² = 0.04). Notably, this interaction also emerged as significant in Experiment 1 in the simulated data (Exp1: F(1, 320) = 6.34, p = 0.01, partial η² = 0.02). Similarly, mixed-effect ANOVAs on the Pseudo SD of predicted individual donation amounts also showed a significant main effect of the phase (Exp1: F(1, 320) = 778.26, p < 0.001, partial η² = 0.71; Exp2: F(1, 341) = 836.19, p < 0.001, partial η² = 0.71; Exp3: F(1, 346) = 762.26, p < 0.001, partial η² = 0.69; Exp4: F(1, 370) = 480.14, p < 0.001, partial η² = 0.57; Fig. S8). In addition, a significant interaction between the phase and the variability of others’ donations was observed across all four experiments (Exp1: F(1, 320) = 4.89, p = 0.03, partial η² = 0.02; Exp2: F(1, 341) = 5.08, p = 0.03, partial η² = 0.02; Exp3: F(1, 346) = 8.75, p = 0.003, partial η² = 0.03; Exp4: F(1, 370) = 7.52, p = 0.006, partial η² = 0.02). Taken together, these posterior predictive check analyses reveal that the winning model provides a close quantitative account of the behavioral patterns observed in the empirical data.

Results S3. Assessing the associations between empathy and the susceptibility to social information

By examining the associations between empathy and the absolute donation changes from the first to the second donation, we found there were no significant association between them in all four experiments (Exp1: Spearman ρ = -0.04, p = 0.51, 95% CI: [-0.15, 0.08]; Exp2: Spearman ρ = - 0.03, p = 0.57, 95% CI: [-0.14, 0.08]; Exp3: Spearman ρ = -0.03, p = 0.58, 95% CI: [-0.14, 0.08]; Exp4: ρ = 0.09, p = 0.086, 95% CI: [-0.02, 0.19]; Fig. S10). Similarly, there were no significant relationships between empathy and the social information weights from the model across all experiments (Exp1: Spearman ρ = -0.01, p = 0.81, 95% CI: [-0.13, 0.10]; Exp2: Spearman ρ = - 0.10, p = 0.062, 95% CI: [-0.21, 0.01]; Exp3: Spearman ρ = -0.06, p = 0.24, 95% CI: [-0.17, 0.05]; Exp4: ρ = 0.09, p = 0.091, 95% CI: [-0.02, 0.19]; Fig. S11). After controlling for age, gender and experimental conditions, no significant effects were observed for the absolute donation changes across all experiments (Exp1: b = -0.000, SE = 0.001, t (318) = -0.29, p = 0.77, 95% CI = [-0.002, 0.001]; Exp2: b = -0.002, SE = 0.005, t (338) = -0.47, p = 0.64, 95% CI = [0.011, 0.007]; Exp3: b = -0.046, SE = 0.041, t (343) = -1.14, p = 0.25, 95% CI = [-0.126, 0.033]; Exp4: b = 0.039, SE = 0.041, t (367) = 0.95, p = 0.34, 95% CI = [-0.041, 0.119]; Table S11). Likewise, the effects of the social information weight on empathy were not significant, after additionally controlling other modeling parameters (Exp1: b = 0.000, SE = 0.001, t (315) = 0.02, p = 0.985, 95% CI = [-0.003, 0.003]; Exp2: b = -0.002, SE =0.001, t (336) = -1.34, p = 0.181, 95% CI = [-0.005, 0.001]; Exp3: b = -0.002, SE = 0.001, t (341) = -1.44, p = 0.151, 95% CI = [-0.005, 0.001]; Exp4: b = 0.002, SE = 0.001, t(365) = 1.47, p = 0.142, 95% CI = [-0.001, 0.005]; Table S12).

In addition, we investigated the association between empathy and social information use in the perceptual task, and also found no significant relationship between the two (Spearman ρ = 0.07, p = 0.18, 95% CI: [-0.03, 0.17]; Fig. S15). This effect remained null after controlling for age and gender (b = 0.001, SE = 0.001, t (365) = 1.15, p = 0.251, 95% CI = [-0.001, 0.003]; Table S14).

Results S4. Assessing the associations between psychopathy, empathy, baseline donation amount and baseline estimation

As preregistered, we also investigated the associations between psychopathy, empathy, and baseline donation levels. We had hypothesized that baseline donation amounts would be positively correlated with empathy, but negatively correlated with psychopathy. The results revealed a significant negative association between psychopathy and the baseline donation amount in Experiment 1-3, but not significant (and in the other direction) in Experiment 4 (Exp1: Spearman ρ = -0.15, p = 0.006, 95% CI: [-0.26, -0.04]; Exp2: Spearman ρ = -0.12, p = 0.024, 95% CI: [-0.23, -0.01]; Exp3: Spearman ρ = -0.11, p = 0.038, 95% CI: [-0.22, 0]; Exp4: ρ = 0.09, p = 0.07, 95% CI: [-0.01, 0.20]; Fig. S12). In contrast, empathy was positively associated with the baseline donation amounts in Experiment 1, 3 and 4, but not significant in Experiment 2 (Exp1: Spearman ρ = 0.17,p = 0.002, 95% CI: [0.06, 0.28]; Exp2: Spearman ρ = 0.02,p = 0.71, 95% CI: [-0.09, 0.13]; Exp3: Spearman ρ = 0.13, p = 0.014, 95% CI: [0.02, 0.24]; Exp4: ρ = 0.22,p < 0.001, 95% CI: [0.12, 0.32]; Fig. S13).

In the perceptual-based social influence task, and as predicted (https://osf.io/ys9zw/overview), psychopathy was not associated with initial estimates of the number of presented animals (Spearman ρ = -0.03, p = 0.51, 95% CI: [-0.14, 0.07]; Fig. S14). Similarly, empathy showed no significant association with participants’ initial estimates (Spearman ρ = 0.06, p = 0.23, 95% CI: [-0.04, 0.17]; Fig. S16).

Supplementary Figures

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Figure S6.

Figure S7.

Figure S8.

Figure S9.

Figure S10.

Figure S11.

Figure S12.

Figure S13.

Figure S14.

Figure S15.

Figure S16.

Supplementary Tables

Table S1.

Table S2.

Table S3.

Table S4.

Table S5.

Table S6.

Table S7.

Table S8.

Table S9.

Table S10.

Table S11.

Table S12.

Table S13.

Table S14.

Data availability

All code and data used to generate results and figures in the paper will be made available upon publication at OSF.

Acknowledgements

We would like to thank Dr. Vladas Griskevicius for his valuable discussions on experimental design. We also want to acknowledge that the Prolific pilot data collection was supported by a Research Grant from the Council of Graduate Students at the University of Minnesota, awarded to Tao Jin. ChatGPT (GPT-5.3, OpenAI) was used for language editing and wording refinement during the manuscript preparation. All outputs were carefully reviewed by the authors. All scientific content, interpretation and final conclusions were performed solely by the authors.

Additional information

Funding

University of Minnesota (UMN) (Internal Funding)

• Iris Vilares

University of Minnesota (UMN) (Research Grant from the Council of Graduate Students)

• Tao Jin