1 Introduction

From collaborating on school projects to managing household responsibilities, humans often engage in cooperative acts that require personal costs (e.g., time, effort, money) for mutual benefit (Tomasello et al., 2005). Such cooperation is central to social life and has shaped human evolution (Nowak, 2006; West et al., 2021). Reciprocity theory views cooperation through the lens of a cost-benefit framework, where people cooperate because the future benefits outweigh the present costs (Nowak & Sigmund, 2005; Panchanathan & Boyd, 2004). It is widely known that betrayal, as an emotional cost in cooperation, provokes negative emotions and promotes punishment (Fehr & Gächter, 2000). While much prior research has only focused on “betrayal aversion”, less attention has been paid to the financial loss from unmet expectations during social interactions (Bohnet et al., 2008; Fehr & Fischbacher, 2003). Indeed, mutual cooperation maximizes social welfare, whereas betrayal benefits the trustee but comes at the trustor’s expense in the Trust Game (Joyce et al., 1995). Here, we compare how “betrayal aversion” and “loss aversion” (i.e., avoiding cooperation to avoid incurring financial losses due to betrayal) differently shape cooperation. Understanding these distinctions is critical for linking computational and neural explanations of cooperative behavior (Krakauer et al., 2017; Lockwood et al., 2020).

Individuals are more sensitive to potential losses than to equivalent gains, a phenomenon known as loss aversion (Kahneman & Tversky, 1979), which explains why investors typically avoid placing all their eggs in one basket. However, risk aversion alone may not fully account for willingness to take risks when the uncertainty arises from another person’s actions (i.e., social risk) rather than nature (i.e., financial risk). Numerous studies demonstrate that betrayal imposes an additional emotional cost beyond monetary loss (Aimone & Houser, 2012). In these studies, participants faced two distinct scenarios: one in which they could keep their money or triple it by giving it to someone else, who might either split it fairly or return only a small portion (less than the original amount), and another involving a lottery-like scenario with identical payoffs, where the return outcome was determined purely by chance. Analyses compared the minimum probability at which people accepted risks between social and non-social domains. A consistent finding is that participants are less willing to take risks in the trust game than in the lottery game (Bohnet et al., 2008; Bohnet & Zeckhauser, 2004), an effect termed betrayal aversion (Bohnet & Zeckhauser, 2004). This aversion is modulated by a variety of factors, including individual characteristics such as gender (Croson & Gneezy, 2009; Finkel et al., 2002) and broad personality traits (Komiya & Mifune, 2015; Thielmann & Hilbig, 2015; Yamagishi, 2011), as well as socio-environmental factors like social status (Hong & Bohnet, 2007), and geographic region (Bohnet et al., 2010). Additionally, psychophysiological factors, such as emotional states (Kugler et al., 2010) and oxytocin (Kosfeld et al., 2005), also shape betrayal aversion. Neuroimaging studies reveal that the differences between betrayal costs and loss costs are tracked by neural activity in the amygdala (van Honk et al., 2013) and insula (Aimone et al., 2014). These findings strongly suggest that social risk and financial risk have some fundamental distinction (Fehr, 2009a; Lauharatanahirun et al., 2012).

While existing studies have focused on how people choose to cooperate when faced with social and financial risks, they have largely overlooked an essential aspect of betrayal cost: the financial loss inherent in betrayal. Specifically, when investors decide to trust trustees, who may either reciprocate or betray, betrayal simultaneously inflicts both emotional harm and financial loss on the trustor (Aimone & Houser, 2011). However, prior studies have treated betrayal aversion and loss aversion as separate phenomena by examining them under distinct uncertain conditions. This methodological limitation leaves unresolved whether the observed reluctance to cooperate stems from the psychological impact of betrayal itself or the direct financial loss it causes, even within the same cooperative context. To address this, we constructed a value-based computational model defining the subjective value of cooperative versus non-cooperative choice, allowing us to dissociate betrayal from loss. We assumed that the best predictive logistic model would incorporate both betrayal- and loss-related components. Moreover, due to the poor temporal resolution of fMRI, previous studies may have conflated neural activities associated with cognitive processes that occur closely in time but are psychologically distinguishable. Indeed, social decision-making can be decomposed into different stages, including early automatic information processing and late-stage elaborative cognitive evaluation, each with distinct ERP components (Fan & Han, 2008; Wang et al., 2023; Yu et al., 2022).

To address these challenges, we employed the event-related potential (ERPs), which provide excellent temporal resolution to characterize the neural dynamics underlying distinct cognitive processes engaged in social decision-making (Luck, 2014). Cooperative choices are mainly associated with two ERP components: the P3 and the late positive potential (LPP). The P3 is a parietally distributed positive deflection that peaks around 350–500 ms after stimulus onset, which reflects a rapid encoding of motivational significance in prosocial actions (Li et al., 2020; Wang et al., 2017), with larger P3 amplitudes signifying greater prosocial motivation (Chiu Loke et al., 2011; Li et al., 2023; Ma et al., 2011). Using a donation task, Carlson et al. (2016) found that that high- versus low-empathy targets increased P3 under intuitive but not reflective decision-making. Directly following the P3, the LPP is a sustained positive deflection over centroparietal areas following the stimulus onset (Hajcak et al., 2009). Unlike the P3, the LPP is thought to reflect the extended, elaborative emotion and motivation processes (Glazer et al., 2018) and might be sensitive to social information evaluation. Crucially, mounting evidence suggests that prosocial actions may stem from fast, automatic, and intuitive processes, while reflective control reflective control can reduce them (Rand et al., 2012; Righetti et al.; Zaki & Mitchell, 2013). Together, these findings illustrate the possibility that the P300 may provide an early motivational signal that fosters intuitive prosocial behavior, while LPP may represent a later, more elaborate evaluation that promotes reflexive prosocial actions. However, previous ERP research has never considered the joint impact of betrayal and loss aversion on cooperation.

To assess the motivations for cooperation and provide computational and neural accounts of betrayal and loss, we combined computational modelling and ERP with two decision-making tasks: a risk task and a cooperation task. They shared identical probabilities (10%-90%) and principal amounts (¥0.2-¥1.0). Participants could opt for the risky choice, which offered the possibility of doubling the principal but also carried the risk of losing half of the principal, or choose the safe option to retain their initial stake. The primary distinction between the tasks lies in the social context: in the cooperation task, a successful outcome resulted in mutual gain with a partner, while failure meant the partner took half of the participant’s principal. In the risk task, losses were solely impersonal. Based on previous studies, we hypothesized that the early P3 reflects intuitive prosocial responses, encoding betrayal aversion, with larger P3 linked to lower betrayal aversion. In contrast, late LPP integrates betrayal and loss aversion after more deliberate evaluation, shown by a more positive LPP as betrayal aversion decreased and loss aversion increased. To provide a comprehensive understanding of cooperation, we also examined participants’ decision-making tendencies across the tasks. We hypothesized that participants would be less willing to take risks facing social risk than financial risk.

2 Results

3.1 Individuals are also less willing to take risks in the cooperation task than in the risk task

As a manipulation check, we examined participants’ rating data as a function of task type and outcome feedback with a series of repeated-measures analyses of variance (ANOVAs). 42 participants were retained for the final analysis due to data recording issues. As shown in Figure 1E, the cooperation task was perceived as less pleasant than the risk task, reflected by a significant main effect of task type, F (1, 41) = 297.01, p < 0.001, ηp² = 0.88. As expected, we also found a significant interaction between task type and outcome feedback, F (1, 41) = 16.75, p < 0.001, ηp² = 0.29. Post hoc comparisons revealed that the cooperation task elicited greater happiness than the risk task in response to positive feedback (8.29 ± 0.17 vs. 7.76 ± 0.19, p = 0.005, Cohen’s d = 0.30). Similarly, the cooperation task elicited greater unhappiness than the risk task with negative feedback (3.12 ± 0.22 vs. 3.83 ± 0.20, p = 0.002, Cohen’s d = 0.41). Together, our rating data suggest that the manipulation of the cooperation task was effective.

Figure 1.

We fitted decision-time data using a linear mixed-effects regression with probability, principal, task type, and their interactions as predictors. As illustrated in Figure 1A–C, decision time decreased as probability (b = -21.51, p < 0.001) and principal magnitude (b = -26.02, p < 0.001) increased. Additionally, a significant interaction between probability and principal magnitude was observed (b = 12.97, p < 0.001). Simple slopes analyses revealed that decision time became shorter as probability level increased at low principal magnitude (M –1SD: b = - 69.00, 95% CI = [-87.50, -50.45], p < 0.001), but not at high principal magnitude (M + 1SD: b = -17.10, 95% CI = [-35.60, 1.43], p = 0.071). Decisions were also faster in the cooperation task than the risk task (b = -73.89, p < 0.001). We also observed significant interactions between probability and task type (b = 26.55, p < 0.001), as well as principal and task type (b = -13.33, p = 0.046). Simple slopes analyses revealed that increased probability level decreased the decision time in the risk task (b = -34.78, 95% CI = [-44.00, -25.53], p < 0.001), but not in the cooperation task (b = -8.24, 95% CI = [-17.5, 1.02], p = 0.081). However, increased principal magnitude decreased the decision time more pronouncedly in the cooperation task (b = -32.70, 95% CI = [-41.90, -23.40], p < 0.001) than in the risk task (b = -19.40, 95% CI = [-28.60, - 10.10], p < 0.001). Full regression estimates are shown in Supplementary Table S1.

Participants’ decision preference was fitted by a mixed-effects logistic regression with probability, principal, task type, and their interactions as predictors. As shown in Figure 2D–E, risk-taking increased as the probability level increased (b = 20.18, p < 0.001) and principal magnitude decreased (b = 0.66, p < 0.001). These effects were further qualified by a significant interaction between probability and principal (b = 1.36, p < 0.001). Simple slopes analyses revealed that participants were more willing to take risks as probability level increased at high principal magnitude (M + 1SD: b = 6.63, 95% CI = [6.00, 7.26], p < 0.001) than at low reward magnitude (M – 1SD: b = 5.39, 95% CI = [4.77, 6.01], p < 0.001). Moreover, they exhibited lower risk-taking in the cooperation task than the risk task (b = 0.44, p < 0.001), further modified by task type × probability (b = 0.64, p < 0.001) and task type × principal interaction (b = 0.90, p < 0.001). Simple slopes analyses revealed that the inhibition effect of the principal on willingness to take risks was more pronounced during the cooperation task (b = -0.47, 95% CI = [-0.53, -0.41], p < 0.001) compared to the risk task (b = -0.37, 95% CI = [-0.44, -0.30], p < 0.001). Similarly, the facilitation effect of probability on willingness to take risks was less pronounced during the cooperation task (b = 2.78, 95% CI = [2.47, 3.09], p < 0.001) compared to the risk task (b = 3.23, 95% CI = [2.91, 3.54], p < 0.001).

Figure 2.

The probability × task type interaction was further qualified by a significant three-way interaction among probability, principal, and task type (b = 0.82, p = 0.002). Post hoc simple slope analyses (Figure 1F) showed that the facilitative effect of probability on risk-taking was attenuated in the cooperation task than the risk task, particularly at high principal magnitude (risk: b = 3.63, 95% CI = [3.29, 3.97], p < 0.001; cooperation: b = 2.99, 95% CI = [2.67, 3.32], p < 0.001) and to a lesser extent at low principal magnitude (risk: b = 2.82, 95% CI = [2.50, 3.14], p < 0.001; cooperation: b = 2.57, 95% CI = [2.26, 2.89], p < 0.001). Full regression estimates are shown in Supplementary Table S1. Overall, participants were less motivated to take risks in the cooperation task than the risk task, as evidenced by lower accepted probabilities and longer decision times in the cooperation task.

3.2 Betrayal aversion suppresses cooperation willingness to a higher extent than loss aversion

We constructed a series of model to test effects of loss aversion and betrayal aversion on cooperation, including (a) a model with linear formulations of utilities for gain and loss values (Model 1); (b) a model that additionally incorporate betrayal aversion into the loss utility (Model 2); (c) a model that assumes participants consider other’s value in addition to gain and loss values (Model 3); and (d) models with task-specific parameters in the above three models (Models 4, 5, and 6) (see Method for more details). Model comparison decisively favored Model 3 (AIC = 238.8), which incorporated CPT, betrayal aversion, and prosocial utility without task-specific parameters (Figure 2A). Parameter recovery analysis showed the parameters of M3 were well identifiable (α_gain = 87%, α_loss = 98%, γ = 99%, λ = 90%, β = 92%, κ = 83%, ω = 88%; Figure 2B and C; details of model comparison and parameter recovery see Supplementary Table S2 & S3).

To examine how betrayal aversion and loss aversion influence cooperation behavior, we fitted decision time using a linear mixed-effects regression and decision preference using a mixed-effects logistic regression with a binomial link. As shown in Figure 2D, both loss aversion and betrayal aversion significantly predicted cooperation, such that participants cooperated less as loss aversion (b = 2.22, p < 0.001) and betrayal aversion (b = 6.96, p < 0.001) increased. Moreover, we employed β coefficient tests to compare the predictive power of loss and betrayal aversion on cooperation willingness, revealing that betrayal aversion had a stronger impact on cooperation than loss aversion (z = -6.97, p < 0.001). Both aversions also robustly predicted decision times, with decision time becoming shorter as loss aversion increased (b = 69.97, p < 0.001) and betrayal aversion decreased (b = -94.04, p < 0.001), again with betrayal aversion showing greater predictive capacity (z = 7.31, p < 0.001). Moreover, as shown in Figure 2E, a greater preference for others’ benefits was associated with a higher betrayal aversion (r = -0.50, p < 0.001), but not with loss aversion (r = 0.01, p = 0.997), indicating that individual differences influence cooperative tendencies. Full regression estimates are provided in Supplementary Table S4. Together, our computational modeling indicates that betrayal aversion has a stronger suppressive effect on cooperative choices than loss aversion, reflected in lower cooperation willingness and longer decision times.

3.3 Betrayal information was encoded during the early P3 period, whereas loss information was further integrated during the late LPP period

The P3 and LPP components were observed as a relative positivity over centroparietal areas (Figure 3). Betrayal and loss parametric were derived from the winning model (Model 3). To investigate their effects on neural dynamics underlying cooperation, we fitted the amplitude of the P3 and LPP using a linear mixed-effect regression with z-scored betrayal and loss parametric as predictors. As shown in Figure 3, the P3 became less positive as betrayal aversion increased (b = 0.44, p = 0.017), but was unaffected by loss aversion (b = -0.09, p = 0.642), indicating a betrayal inhibition effect. For the LPP, both betrayal inhibition and loss inhibition effects were observed, with reduced positivity as betrayal aversion (b = 0.71, p < 0.001) and loss aversion (b = -0.48, p = 0.009) increased. Further β coefficient tests showed comparable predictive power of betrayal and loss on LPP amplitude (z = -0.85, p = 0.198). Full regression estimates are shown in Supplementary Table S5. Together, our ERP data suggest that betrayal and loss information were processed serially, such that betrayal information was encoded during the P3 periods, whereas loss information was not tracked until the time window of the late positive potential.

Figure 3.

3 Discussion

Betrayal suppresses cooperation and trust, yet the associated financial losses it incurs are often overlooked, complicating the reasons behind cooperative behavior. Here, we explore how betrayal and its financial consequences influence cooperation alongside their underlying neural mechanisms, addressing a gap in previous research. We found that betrayal exerts a significantly stronger suppression on cooperation than loss. Moreover, neural processing unfolded sequentially: betrayal modulated early-stage automatic processing indexed by the P3, while loss appeared in later evaluative stages captured by the LPP.

In our study, participants reported a stronger preference for positive feedback and greater aversion to negative feedback during the cooperation task, suggesting that betrayal imposes an additional emotional cost beyond monetary loss. They were also less willing to take risks as the cooperation risk increased, especially when the principal magnitude was high, compared to the financial risk. Moreover, in the cooperation task, decision times were shorter and driven more by principal amounts rather than probability levels, indicating that participants prioritized potential gains and losses and adopted a generally cautious approach, regardless of the cooperative probability. These findings are largely consistent with previous studies, which compared individuals’ willingness to take risks when the outcome depends on another player’s trustworthiness, versus situations where the outcome is determined by a chance with identical odds and payoffs, demonstrating a stronger avoidance motivation in social contexts (Bohnet et al., 2008; Bohnet et al., 2010; Bohnet & Zeckhauser, 2004). However, these studies overlook that betrayal simultaneously incurs emotional and financial losses.

By employing computational modeling to separately examine how betrayal and its associated loss parameters influence cooperation, our most important finding is a greater substantial inhibitory effect of betrayal on cooperation than loss. Participants exhibited lower cooperation rates and longer decision times as both loss aversion and betrayal aversion increased. Notably, betrayal aversion had a stronger impact on choice or decision time. Several theoretical frameworks can help interpret this effect. First, individuals care about others’ payoffs in cooperative situations (Fehr, 2009b). In our study, when partners betrayed by taking half of the investor’s money, participants are less likely to accept the social risk compared to the financial risks. This possibility was further supported by our modeling results regarding the correlation between attention to others, betrayal, and loss aversion. Specifically, those who focused more on others’ benefits show higher betrayal aversion, but not loss aversion. Second, people care not only about outcomes but also the intentions behind them (Charness & Rabin, 2002; Rabin, 1993). Trust builds honor and integrity, promoting cooperation, while betrayal erodes these qualities, discouraging future cooperation. Third, betrayal often inflicts emotional harm, such as damaged trust, which can outweigh material loss. Aimone et al. (2014) found stronger activation in the anterior insular cortex when playing with a human versus a computer, indicating heightened negative emotions (Kuhnen & Knutson, 2005; Wu et al., 2012), supporting betrayal aversion as a desire to avoid negative emotions. Finally, people perceive risks from others’ actions as less controllable than natural risks (Slovic, 1987). For example, shareholders would prefer a 1% chance of losing half their value due to a natural disaster than a slightly smaller chance from betrayal by corporate executives.

Interestingly, we observed distinct neural mechanisms for betrayal and loss. Specifically, the early P3 amplitude decreased with increasing betrayal aversion, while no significant change was observed in response to loss aversion. In contrast, the LPP amplitude decreased as both betrayal and loss aversion increased. These findings suggest that betrayal and loss information are processed sequentially, with betrayal being encoded during the P3 period, followed by further integration of loss information during the LPP period. Cooperation is often considered an intuitive behavior (Gao et al., 2020; Koch et al., 2020; Levine et al., 2018; Rand et al., 2012; Shank et al., 2019), which is indexed by P3, a neural signal that rapidly encodes prosocial motivation (Li et al., 2020; Wang et al., 2017). Betrayal, a key factor inhibiting prosocial behavior (Fehr & Gächter, 2000), is therefore prioritized for encoding at the early stages of cooperation decision-making. According to a dual-process framework, intuition is fast, automatic, and effortless, while reflection is often slow, deliberate, and characterized by the rejection of emotional influence. (Chaiken & Trope, 1999; Frederick, 2005; Kahneman, 2011; Plessner et al., 2008; Sloman, 1996; Stanovich & West, 1998). Hence, during the later phase of decision-making, participants further engage in a trade-off between costs and benefits associated with specific actions. This process is indexed by the LPP, a neural marker reflecting sustained and elaborative information processing (Glazer et al., 2018). In this stage, loss-related information is integrated into the decision-making process.

P300 amplitude may serve as a neural marker of aversion to defection in cooperative interactions. Specifically, the P300 amplitude increases when individuals are faced with decisions that challenge cooperation, particularly when they are prompted by intuitive, emotionally-driven responses. This suggests that P300 could signal the underlying motivational processes that discourage defection and promote prosocial behavior, especially when individuals feel a social or emotional connection to others involved in the interaction. In this context, the P300 response not only provides insight into how cooperative behaviors are initiated but also offers a predictive measure of one’s commitment to maintaining cooperation under varying conditions of social engagement.

A limitation of this study is the ecological validity of the cooperative scenario. Participants were informed that they would cooperate with others, but some (N=6) expressed doubts about the authenticity of this interaction in a post-experiment check. Although this did not affect the results, future studies could investigate neural synchronization during multi-player cooperation to better understand how co-players synchronize their neural activity in such contexts (Jiang et al.; Yang et al., 2020; Zhang et al., 2023). Additionally, future studies could provide evidence to distinguish the aversion to being personally betrayed and witnessing another’s betrayal. This approach will help determine whether a betrayal has a stronger inhibitory effect and is processed with priority from a third-party perspective, where the outcome is unrelated to oneself.

In conclusion, this study demonstrates that betrayal has a stronger suppressive effect on cooperation compared to the associated cost of losses. Additionally, this study provides preliminary evidence that the neural dynamics involved in these two forms of aversion occur sequentially, with betrayal aversion influencing early-stage decision-making and loss aversion affecting later-stage evaluations. These findings enhance our understanding of how emotional and financial considerations influence cooperative behavior and provide important insights into the neural mechanisms that underlie cooperative decision-making.

4 Materials and methods

All data and code used for this study are available on OSF at https://osf.io/zw2ra/. This study was not preregistered.

4.1 Participants

Fifty right-handed university students were recruited via local advertisements for this study. One participant was excluded from data analysis due to a misunderstanding of instructions. The final sample thus consisted of 49 participants (24 females; M = 22.00 years, standard deviation [SD] = 2.03). We performed a sensitivity analysis using the simr v1.0.6 package(Green & MacLeod, 2016) (Green & MacLeod, 2016) to compare the regression weight for each effect of interest with the smallest detectable effect size at a power of 80% based on the current sample. The results showed that most of the significant effects observed were larger than the smallest detectable effect, suggesting that our sample size provided adequate statistical power. All participants had normal or corrected-to-normal vision as determined by self-report and no psychiatric or neurological disorders. Each received a bonus of ¥70–¥80 based on their task performance. This study was approved by the Institutional Review Board of Sun Yat-sen University.

4.2 Procedure

Upon arrival at the lab, participants completed a probabilistic risk task and a single-shot cooperation task while their EEG was recorded. After the EEG tasks, participants rated their perceived happiness on a 9-point Likert scale (ranging from 1 = not at all to 9 = very much), based on the outcomes (gain or loss) associated with each decision.

The risk task

This task was designed to assess neural correlates of loss aversion. On each trial (Figure 4A), participants first viewed a probability pie chart for 1500 ms, followed by a jittered interval (1200–1500 ms). The pie chart illustrated probability levels (10%, 30%, 50%, 70%, or 90%, corresponding to 1, 3, 5, 7, or 9 segments of the chart), indicating the likelihood of doubling the principal versus losing half of it. Afterward, a number appeared for 1500 ms, showing the principal amount (¥0.2, ¥0.4, ¥0.6, ¥0.8, or ¥1.0, corresponding to 1–5 principal levels). The five probability levels were fully crossed with the five principal levels, resulting in 25 unique combinations. Following another jittered interval (900–1100 ms), participants entered the choice phase, choosing between the “reject” option (R), which allowed them to keep the current principal and proceed to the next trial, and the “accept” option (A), which offered a chance to double the principal with the given probability or lose half in case of failure.

Figure 4.

Participants had up to 4000 ms to decide using their left or right index finger. The chosen option was highlighted with a green border for 300 ms. Failure to respond within the time limit resulted in ¥0 and a 1000-ms warning. After another jittered interval (1200–1500 ms), an outcome feedback stimulus was presented for 1000 ms: a tick denoted gain, a cross denoted loss, followed by a jittered interval of 900–1100 ms. Subsequently, a reward feedback stimulus was shown for 1000 ms, indicating the final payoff. Each trial ended with an interval varying between 900 and 1100 ms.

The cooperation task

This task was designed to measure participants’ neural responses to both loss aversion and defect aversion, which employed a similar structure but incorporated social decision-making elements. One key distinction was that each trial involved an anonymous co-player, who was distinct from the ones in any other trials and would not reappear in any subsequent trial. Another key difference was the presence of a probability cue reflecting the likelihood that the co-player would cooperate. Participants were informed that the probabilities as derived from prior behavioral data, but in reality, they were randomly generated. Critically, unlike the risk task, where outcomes depended solely on chance, the cooperation task introduced interpersonal contingencies. During the choice phase, participants decided whether to cooperate (C) or not (N). Feedback then revealed the co-player’s decision (a tick for cooperation, a cross for defection), followed by a reward screen showing the monetary outcome for both players. Mutual cooperation resulted in the principal being doubled and split equally. If the participant cooperated but the co-player defected, the participant incurred a trust cost by forfeiting half of their principal, which was transferred to the co-player in addition to the co-player’s own principal. If the participant chose not to cooperate, the principal remained unchanged regardless of the co-player’s choice. The risk task and cooperation decision-making task each consisted of 200 trials, divided into eight blocks of 25 trials, with a self-determined break between blocks. The task order was counterbalanced across participants. To maintain attention, each task included 10 catch trials in which participants were required to confirm the probability and principal levels for the current trial. Before the experiment, participants completed 7 practice trials for each task for familiarization. After the task, participants were asked to report whether they believed they were playing with the third-party players. Three subjects had an accuracy rate of less than 60% on catch trials, and six subjects expressed doubts about the cooperation manipulation. However, since this did not affect the results, they were included in the analysis.

4.3 EEG recording and processing

EEG data were recorded using 64 Ag/AgCl channels placed on an elastic cap based on the international 10–20 system. Two additional channels were positioned on the left and right mastoids. The reference electrode was located at Cz. Horizontal and vertical electrooculograms were recorded from two pairs of channels over the external canthi of each eye and the left suborbital and supraorbital ridges, respectively. EEG signals were amplified using a Neuroscan SynAmps² amplifier with a low-pass filter of 100 Hz in DC acquisition mode and digitized at a rate of 1000 samples per second. Channel impedances were maintained below 10 KΩ.

The EEG data were analyzed using EEGLAB v2021.0 (Delorme & Makeig, 2004) and ERPLAB v8.10 (Lopez-Calderon & Luck, 2014) toolboxes in MATLAB 2020b (MathWorks, US). The signals were rereferenced to the average of the left and right mastoids and filtered with a bandpass of 0.1–35 Hz using a zero phase-shift Butterworth filter (12 dB/octave roll-off). Channels with poor quality or excessive noise were interpolated using the spherical interpolation algorithm, and portions of EEG containing extreme voltage offsets or break periods were removed. Ocular artifacts were removed using an infomax independent component analysis on continuous EEG with the help of the ICLabel algorithm (Pion-Tonachini et al., 2019). Epochs were then extracted from -200 to 1000 ms relative to principal stimulus onset, with the prestimulus average activity as the baseline. An automatic artifact detection algorithm was applied to remove epochs with a voltage difference exceeding 50 μV between sample points or 200 μV within a trial, a maximum voltage difference less than 0.5 μV within 100-ms intervals, or a slow voltage drift with a slope greater than ± 100 μV. On average, 96.30% of trials were retained for statistical analysis. Measurement parameters were determined by the grand-averaged ERP waveforms and topographic maps collapsed across all conditions (Luck & Gaspelin, 2017). Specifically, the single-trial P300 amplitude and LPP amplitude were measured as the mean activity from 400 to 500 ms and 600 to 800ms, respectively, after the onset of the principal cues over centroparietal areas (P1, Pz, P2). These single-trial data were exported into R v4.2.2 for statistical analyses.

4.4 Data analysis

Statistical analysis

Our key statistical analyses were based on mixed-effects regression models with random intercepts and slopes (unstructured covariance matrix), implemented in the lme4 package v1.1.31 (Bates, Maechler, et al., 2015). For decision time and choice data, we separately used a linear mixed-effects regression model and a mixed-effects logistic regression model with a binomial link function. Both models included probability, principal, task type, and their interactions as predictors. For all models, categorical regressors (task type: -0.5 for risk and +0.5 for cooperation) were contrast-coded, and continuous regressors (probability and principal) were z-scored before inclusion in the regression models. We determined random effects for each model using singular value decomposition to report the maximal possible random effects structure (Barr, 2013; Bates, Kliegl, et al., 2015). Follow-up pairwise comparisons of significant interactions were tested on estimated marginal means using the emmeans package v1.8.3 (Lenth, 2022). To examine the relationship between behavioral indices of loss aversion (λ) and defect aversion (κ) and neural responses to principal cues, both λ and κ were entered as fixed effects in separate P300 and LPP models. All model predictors were scaled (e.g., λ/sqrt(sum(λ^2)/length(λ)-1)) without mean-centering to ensure comparability of β coefficients. We analyze post-experimental rating data using a repeated-measures analysis of variance (ANOVA), with feedback (good, bad) and task type (risk, cooperation) as within-subjects factors. We excluded trials with no responses (0.15%) in the risk and cooperation task from statistical analyses.

Computational modeling

To quantitatively dissociate the contributions of loss aversion and betrayal aversion to cooperative decision making, we employed a computational framework grounded in Cumulative Prospect Theory (CPT; Tversky & Kahneman, 1992). This approach models risky decisions through two core transformations: a value function that asymmetrically weights losses versus gains, and a probability weighting function that nonlinearly distorts objective probability. A family of seven models (M0-M6) was systematically constructed to incorporate socially specific factors and test hypotheses about how financial losses and betrayal aversion independently affect cooperation choices. The simplest model (M0) served as a random choice baseline. Model 1 was adapted from the cumulative prospect theory, and defined subjective value (SV) as the net expected utility of choosing investment or cooperation. Specifically, for a given option with principal amount X and success probability p, the SV was computed as

Where the net utility of success ΔX_success = X and net utility of failure ΔX_failure = −0.5 ∗ X as the setting of the experiment. The subjective utility function V(x) was defined as

Here, the parameter α_gain and α_loss (0 < α < 1) represents the diminishing sensitivity to increasing magnitudes of gains and losses, respectively; λ quantifies financial loss aversion, λ ≥ 1 indicates that the reduced utility of losses is greater than the utility of gains at the same magnitude.

Besides, the subjective probability weighting function w(p) was defined as:

Parameter γ (γ < 1) governing the curvature and sensitivity of probability for the outcomes. For the simplicity of the model, we estimated a single γ parameter across gains and losses. Lower values on γ indicate greater curvature and lower sensitivity to probabilities.

Building on M1, Model 2 introduced betrayal aversion as an additive emotional cost term independent of loss aversion. The SV of M2 was computed as

Here, k (k > 0) captures betrayal aversion independent of the monetary loss magnitude, reflecting the hypothesis that social betrayal carries intrinsic disutility beyond monetary consequences.

Model 3 further incorporated explicit valuation of the co-player rewards to capture prosocial preferences independent of self-interest.

Here, ω is a scaling factor, and the term inside the parentheses reflects the expected reward for others set by experiment. Then, the SV of M3 was computed as

The derived SV of the option was transformed into choice probability through a softmax function incorporating a stochasticity parameter β:

Where β (β ≥ 0) represents the slope of the logistic function and reflects the sensitivity to SV differences between options. The lower β is, the greater the stochasticity of the choices; with β = 0, choices are random (Collins & Shenhav, 2022).

To account for potential task-dependent variations, we tested three additional variants (M4-M6), allowing the sensitivity of gain α_gain, the sensitivity of probability γ, and the stochasticity parameter β in M1-M3 vary across tasks.

We estimated the parameters of each model for each participant, using maximum likelihood estimation implemented by fmincon function in MATLAB, with multiple initialization points to avoid local minima. Model comparison employed the Akaike Information Criterion (AIC) to balance model fitting and complexity. A lower AIC value indicates a better-fitting and concise model. Moreover, the parameters’ identifiability was validated through recovery analyses. For each model, 500 plausible parameter sets were randomly sampled from uniform distributions bounded by empirically informed ranges. Using these ground-truth parameters, we simulated trial-wise choice data (500 trials/set) under experimental conditions matching the original task design. The model was subsequently refitted to the simulated datasets using 20 randomized initialization points per recovery to mitigate local minimum convergence. Parameter recovery similarity was quantified through Pearson correlations between true and estimated parameter values across all simulations.

Supplementary Materials

Table S1.

Table S2.

Table S3.

Table S4.

Table S5.

Acknowledgements

This work was supported by the National Natural Science Foundation of China (32171073).

Additional information

Funding

MOST | National Natural Science Foundation of China (NSFC)

• Dingguo Gao