Betrayal is worse than loss during cooperation

  1. Rumeng Tang
  2. Jingbin Tan
  3. Yi Gao
  4. Chen Lin
  5. Jing Gan
  6. Xiaowei Ding
  7. Dingguo Gao

  • Curated by eLife

    eLife logo

    eLife Assessment

    This valuable study examines whether reduced cooperation is driven by betrayal aversion beyond nonsocial loss aversion, using matched social and nonsocial risky decision-making tasks combined with computational modeling and EEG. The authors provide solid empirical evidence that social risk is processed differently from matched nonsocial risk, offering a meaningful contribution to the study of cooperation and decision-making under uncertainty. However, further justification of the computational modeling approach would strengthen some of the conclusions. This work will be of interest to researchers studying social decision-making, cooperation, trust, and the neural and computational mechanisms underlying risk and betrayal aversion.

Reviewed by

Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

Log in to save this article

Abstract

Cooperative behavior is a cornerstone of human interaction. Although both “betrayal aversion” (the affective cost of being betrayed) and “loss aversion” (the financial detriment incurred from betrayal) are established determinants of cooperative behavior, their relative potency remains undetermined. Here, we investigated these effects by integrating computational modeling and event-related potential (ERP) techniques. In two tasks involving risk and cooperation, participants decided whether to take financial risks or to cooperate under possible betrayal. Our results showed that betrayal aversion had a stronger effect on reducing cooperation compared to loss aversion. Furthermore, ERP data demonstrated sequential processing: betrayal was encoded early in decision-making, reflected by increased P3 with weaker betrayal aversion, whereas loss aversion manifested later, marked by increased LPP. By dissociating the contributions of betrayal and loss, our finding provides novel insights into the cognitive and neural mechanisms underlying cooperative behavior.

Article activity feed

  1. eLife

    eLife Assessment

    This valuable study examines whether reduced cooperation is driven by betrayal aversion beyond nonsocial loss aversion, using matched social and nonsocial risky decision-making tasks combined with computational modeling and EEG. The authors provide solid empirical evidence that social risk is processed differently from matched nonsocial risk, offering a meaningful contribution to the study of cooperation and decision-making under uncertainty. However, further justification of the computational modeling approach would strengthen some of the conclusions. This work will be of interest to researchers studying social decision-making, cooperation, trust, and the neural and computational mechanisms underlying risk and betrayal aversion.

  2. eLife

    Reviewer #1 (Public review):

    Summary:

    The non-social task was a classic risky decision-making task with a binary choice between an option with a sure gain and a risky option with a probabilistic gain or loss. In the social task, the sure option was an individual gain (as in the non-social option) and the probabilities in the risky option, which were shown to participants, were framed as probabilities of other previous participants (i.e., "partners") to cooperate or not; a probabilistic gain (when the partner cooperated) also led to a gain of the partner, while a probabilistic loss meant that the partner would receive the amount lost by the participant. This loss was framed as "betrayal." The authors show differences in how probabilities and amounts (of gains/losses) affected choices, RTs, and ERPs (P3 and LPP).

    Strengths:

    Since participants faced decisions with the same individual payoffs in a non-social and a social condition, this setup made it possible to use identical standard analyses for choices, RTs, and ERPS as well as (almost) identical economic models for the two conditions.

    Weaknesses:

    (1) The task does not include many components that are usually considered central for cooperation or "betrayal" and this is not discussed appropriately. At the same time, the "emotional aspects" of the operationalized "betrayal" are not directly assessed.

    a) The standard economic game for cooperation is the prisoner's dilemma, in which participants make independent choices at the same time without getting any explicit information on the cooperation probability of their partner before they make their decisions. Furthermore, most of the time the interactions are repeated. Actually, the trust game as one other frequently used economic game, also includes a back and forth of transfers between the partners. So, here, I am not so convinced by the operationalization of a low cooperation probability, which is shown before the decision, as "betrayal." The authors should motivate and explain their rationale more clearly in reference to such other tasks.

    b) The setup of the task, especially the fake interaction with the fake partners, should be made clearer in the main text (before reporting the results). I would argue for including the task picture in the main text.

    c) In general, I am in favour of taking participants' choice behaviour as the main outcome measure. But given the strong implications of "emotional costs" made by the authors, I would have expected some ratings of "betrayal" on a trial-by-trial basis. I would at least include this as a shortcoming.

    d) Also, given the framing of the study, I would have expected some exploratory analyses regarding individual differences with respect to, e.g., social value orientation, etc. I would at least include this as an outlook.

    (2) The standard statistical analyses could be improved.

    a) It is good that the authors have rather long sections using standard regression analyses. But they are a bit lengthy, and the modelling should be more prominent.

    b) In a couple of places, the authors say something like "this is significant, but that is not." Here, it has been made very clear that the interaction term needs to be looked at. As far as I can see, this has not always been done.

    c) For this binary choice, the difference in expected value (EV) between the sure and the risky options is one crucial comparison. But the authors never take that into account. This difference does not depend on the amount, which the authors dub "principal." That is, the sure option simply has an EV of x, i.e., the amount. The risky option has the EV = p2x + (1-p)0.5x, with p being the probability of gain/cooperation. That is, the two options have the same EV at p=1/3, independent of x. This should be made clear.

    d) Relatedly, RTs should depend on the differences in EV (and not so much on p or on x per se). This can be seen by the more or less quadratic relationship between p and RTs (Fig 1A), with a peak around a p of 1/3.

    e) RTs are often log-transformed. It should be briefly mentioned why this was not done here.

    (3) The modelling evidence is relatively weak. This is my main point.

    a) (Cumulative) prospect theory should be introduced.

    b) The models seem overly complicated with many free parameters. I would have expected some simpler versions and more comparisons between models that differ in just one parameter.

    - e.g., it is really nice that the authors used a probability weighting function. BTW: Please describe this more clearly in the introduction and in the results. But for this limited range of probabilities, this might be too much.

    - e.g., why directly assume two different exponents in the utility function for gains and losses, and in addition a loss aversion parameter lambda? Only lambda would be a better starting point here.

    c) The differences in AIC (Figure 2A) seem rather minuscule, and the distribution of winning models is not very peaked. I am not convinced that Model 3 is the winning model.

    d) Crucially, and related to the previous points, judging from Fig 2C, the "betrayal" parameter kappa seems to be zero for about half of the participants. The authors should look into this.

    - Would a model just like model 3 but without kappa (i.e., kappa set to zero) perform better? Is this just model 2?

    - How is kappa set in the non-social condition?

    - This massive skew, to say the least, is never discussed.

    - A correlation is definitely not warranted.

    (4) The ERP results seem to me rather superficial. But I am not an EEG expert.

    a) The authors do not seem to look at the outcome phase, which could be interesting for differences in reward/loss processing in the two task versions.

    b) Again, differences in EV seem to be more important from a conceptual point than probabilities or amounts; see my comment 2d.

    c) Also, the authors report ERPs for the two task types separately but do not seem to run proper comparisons between them, see my comment 2b.

    (5) Preregistration: It should be made very clear early on that this study was not preregistered.

    (6) Quality checks: The authors should check if some participants are outliers in terms of the number of missed trials, always choosing the same option, etc. It is notoriously difficult to find good post hoc reasons for excluding participants (one reason why replications and preregistrations are important). In any case, the data quality should be checked and described a bit more.

  3. eLife

    Reviewer #2 (Public review):

    Summary:

    This paper investigates risk and cooperation decisions by integrating computational modeling with event-related potential (ERP) measures. Participants completed two tasks involving financial risk and cooperation under possible betrayal. The comparison between social and non-social decision-making is interesting and potentially valuable. However, the conceptual framing, theoretical grounding, and modeling rationale require substantial clarification.

    Strengths:

    (1) The paper introduces comparable tasks to probe social vs. non-social decision making.

    (2) The authors use a model to identify a psychological distinction and test its validity using neural data.

    Weaknesses:

    (1) Conceptual framing and theoretical clarity

    The primary theoretical contribution of the paper is currently unclear. Specifically, it is not clear what key difference the authors hypothesize between risk and cooperation conditions. This distinction should be grounded in prior literature.

    The manuscript states: "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)." However, the authors do not discuss the substantial literature on the Trust Game, which is used here but not explicitly acknowledged.

    • The original Trust Game framework and behavior in one-shot settings (e.g., Berg et al., 1995).

    • The persistence of cooperation even when defection is economically optimal (e.g., Berg et al., 1995; Fehr & Fischbacher, 2003).

    • The influence of trustworthiness of the partner on cooperation decisions has been previously studied (Ma et al., 2022).

    • Differences between social and non-social decision-making contexts have also been reported with matched tasks (Liu et al., 2024).

    (2) Distinction between constructs (risk, loss aversion, betrayal aversion)

    The introduction introduces multiple related constructs-risk aversion, loss aversion, and betrayal aversion-but does not clearly differentiate them. A theoretically grounded distinction is needed.

    In particular:

    • The manuscript introduces multiple related constructs, or maybe the terms are used interchangeably? The distinction between risk aversion, loss aversion, defection aversion, and betrayal aversion should be clearly defined.

    • Betrayal aversion versus loss aversion is introduced but not clearly differentiated. Importantly, it should be clarified that this distinction is not experimentally manipulated but instead inferred through computational modeling. This point is currently not made explicit, which leads to confusion in the introduction

    • The computational model should be introduced clearly in the introduction. Without explaining how these constructs are operationalized in the model, the framework is difficult to follow.
    The statement "In the risk task, losses were solely impersonal" is also unclear. It seems the authors may mean "personal or non-social" rather than "impersonal" as rewards are always personally relevant.

    (3) Hypotheses and preregistration

    The manuscript would benefit from more theoretical rationale for hypotheses. For example:

    • What is the basis for hypothesizing that financial loss aversion and betrayal aversion independently affect cooperation choices?

    • Why should these constructs be separable and modeled independently?

    • Additionally, the absence of preregistration is a limitation that should be acknowledged even more.

    • Given the flexibility of the modeling approach and number of parameters, this is particularly important.

    • For instance, the rationale for focusing on decision times is also not clearly explained and should be better motivated.

    (4) Computational modeling

    There are several concerns regarding the modeling approach:

    • The choice of model comparison metric should be justified. Why is AIC used rather than BIC, which penalizes model complexity more strongly? This is particularly relevant given the inclusion of additional parameters to capture processes not directly measured by the task.

    • Full model recovery analyses are missing. A full model recovery is necessary to demonstrate that competing models produce distinguishable behavioral patterns. This needs to be shown in order to justify the specificity of the winning model

    • How correlated are the parameters across participants, particularly loss and betrayal parameters?

    • More broadly, it is unclear how well loss aversion and betrayal aversion can be differentiated based on behavior alone. If these constructs are separable, they should predict distinct aspects of behavior.

    (5) ERP analyses

    The ERP results (e.g., P300 and LPP) seem to suggest that betrayal aversion is relevant in both time periods and similarly.

    • Do neural signals differentially reflect betrayal aversion versus loss aversion earlier and later on?

    • Are there significant interaction effects between betrayal and loss aversion for each ERP component?

  4. eLife

    Reviewer #3 (Public review):

    Summary:

    In this study, the authors aim to address two questions. First, do people avoid cooperation primarily because of betrayal aversion beyond loss aversion? Second, can the effects of betrayal aversion and loss aversion be dissociated at the behavioral and neural levels? To address these questions, the authors compared individuals' choices of taking risks in a nonsocial risk task with those in a social cooperation task, with the two tasks matched in success probability and principal amount. They fitted computational models that include betrayal-aversion and loss-aversion terms and related the model parameters to ERP measures. Based on these analyses, the authors concluded that betrayal aversion has a stronger effect on cooperation than loss aversion and that betrayal is encoded earlier than loss in the brain. This is an important research question, and the attempt to combine computational modeling with ERP analysis is valuable. However, the current data analyses may not be able to support all the conclusions the authors made. For instance, the claims concerning the dissociation between betrayal aversion and loss aversion are not yet sufficiently supported by the evidence.

    Strengths:

    (1) The research question is theoretically important. Distinguishing betrayal aversion from loss aversion is important for research on trust, cooperation, and risky decision-making.

    (2) The approach of integrating behavioral measures, self-report ratings, computational modeling, and ERP data is valuable and gives the study significance.

    (3) The behavioral findings are broadly consistent. Participants reported stronger emotional responses in the cooperation task and were less willing to accept risk in the cooperation condition. These findings are generally in line with previous work on betrayal aversion and provide a reasonable manipulation check for the contrast between social and nonsocial risk.

    Weaknesses:

    (1) The manuscript states that the two tasks are matched in probability and principal amount, but the cooperation task additionally introduces partner outcomes, betrayal, and prosocial components. The Methods section states that, in the cooperation task, if both players cooperate, the principal is doubled and then split equally; if the partner betrays, half of the participant's principal is transferred to the partner. The model also includes an expected-other-reward term, namely, V_other=ω[p⋅2X+(1-p)⋅1.5X]. This raises an interpretive concern: if the two tasks differ not only in whether the source of uncertainty is social, but also in partner outcome, intentionality, and potential inequity structure, then the fitted "betrayal aversion" parameter may in fact reflect multiple motives rather than betrayal aversion alone. In the current experimental design, the "betrayal aversion" parameter may not be uniquely interpretable as a pure betrayal-specific construct, and the current evidence is insufficient to support such a specific interpretation.

    (2) Participants were informed that the cooperation probabilities were derived from previous real participants, whereas in fact these probabilities were randomly generated. In addition, six participants explicitly expressed doubts about the authenticity of the social interaction, yet the authors retained these participants with only the brief statement that this "did not affect the results." For such a critical manipulation, this explanation is too brief. I recommend that the authors report robustness analyses excluding skeptical participants. Since six participants reportedly doubted the authenticity of the social interaction, and some participants also performed poorly on the catch trials, it would be important to show whether the main behavioral, modeling, and ERP findings remain after excluding these participants. This is especially important because the manuscript's central interpretation depends on the assumption that the cooperation task was genuinely experienced as social.

    (3) The descriptions of the sample size are inconsistent across sections. The Participants section states that, after excluding one participant for misunderstanding the instructions, the final sample consisted of 49 participants; however, the behavioral results section later states that only 42 participants were included in the final analyses due to recording problems. This discrepancy is important because readers need to know clearly which sample was used for the behavioral analyses, which for the model fitting, and which for the ERP analyses; whether these analyses were conducted on the same participants; and whether the exclusion criteria were consistent across analyses. The manuscript needs a more transparent description of sample size and exclusion criteria.

    (4) The authors need to do more thorough analyses to validate their models. In addition to AIC and parameter recovery, I would encourage the authors to include other model comparison metrics where possible, such as BIC and exceedance probability, as well as model-recovery analyses. The authors should also do model-based simulation analyses to show that the winning model can capture the contextual effects observed in real data.

    (5) The authors should explain the rationales for the choice of ERP time windows and component selection in more detail. The current ERP analyses are time-locked to principal onset, and P3/LPP are extracted from fixed time windows. The authors should explain why this is the most appropriate time-locking point for examining betrayal- and loss-related computations, and why alternative time-locking points, such as probability-cue onset or other key task events, were not used. More importantly, the time windows of P3 and LPP are defined arbitrarily in the current analyses. The authors need to apply a more principled approach to define ERP components. It looks like the P3 and LPP are from the same ERP component in Figure 3.

    (6) The manuscript has several internal inconsistencies in terminology, figure references, and result descriptions. These issues weaken the clarity of the arguments and reduce the readability of the manuscript.

    (7) The authors partially achieved their aims. The study does provide evidence that social risk and nonsocial risk are not treated equivalently, and it also offers a computational framework that is informative for the field. This is an important topic, and the overall approach is promising.

  5. eLife

    Author response:

    We agree that the manuscript would benefit from a more clearly articulated conceptual framing, stronger model validation, more explicit statistical and ERP comparisons, and improved transparency regarding task design, sample inclusion, and preregistration. In the revised manuscript, we plan to address these points through substantial revision of the Introduction and Discussion, along with additional robustness and validation analyses, and more cautious interpretation of the main findings.

    Reviewer #1 raised important points about the framing of the cooperation task, the interpretation of betrayal, the standard statistical analyses, the modelling, and the ERP analyses. In response, we plan to clarify that the present task captures betrayal-related social risk or anticipated partner defection, rather than betrayal in its full interpersonal and emotional sense, and to better motivate this operationalization with reference to the betrayal-aversion and trust-game literature. We will moderate our claims regarding “emotional costs,” incorporate a more explicit task overview and accompanying schematic into the main text, and frame individual differences as a key avenue for future research. In addition, we will streamline the standard behavioral analyses, make the expected-value structure of the task explicit, add EV-based analyses of choice and reaction time, strengthen the ERP analyses, clarify that the study was not preregistered, and provide a complete report of data-quality checks. For the modelling section, a central revision will be to simplify the model structure and refit the models using a Bayesian hierarchical approach.

    Reviewer #2 emphasized the need for stronger theoretical framing and more specific distinctions between related constructs. In the revised manuscript, we will substantially revise the Introduction to better situate the present task in relation to the Trust Game literature and prior work comparing social and non-social decision-making under matched payoff structures. We will also define risk aversion, loss aversion, anticipated partner defection, and betrayal-related aversion more explicitly, and clarify that the distinction between betrayal-related aversion and loss aversion is inferred through computational modelling rather than directly manipulated as separate experimental factors. We also plan to introduce the computational model earlier in the manuscript, clarify how the key constructs are operationalized, replace unclear wording such as “impersonal losses,” strengthen the rationale for our hypotheses, and acknowledge the lack of preregistration more clearly.

    Reviewer #3 highlighted the need to align our conclusions more closely with the current evidence. In the revised manuscript, we will moderate the interpretation of the betrayal-related parameter, acknowledging that the cooperation task differs from the non-social risk task not only in social versus non-social uncertainty, but also in partner outcome, intentionality, and potential inequity structure. We therefore plan to avoid treating this parameter as a pure betrayal-specific construct and to describe it more cautiously as capturing betrayal-related social risk or aversion to anticipated partner defection. We also plan to report robustness analyses excluding participants who expressed doubts about the social interaction, as well as participants with poor catch-trial performance or otherwise low-quality data, and to clarify the sample sizes and exclusion criteria used for behavioral, modelling, and ERP analyses. Finally, we will strengthen model validation and ERP reporting, including broader validation analyses and more cautious interpretation if the evidence for temporal dissociation between betrayal-related aversion and loss aversion proves weaker than currently stated.

    Across these revisions, we also intend to simplify the model structure and use Bayesian hierarchical fitting to strengthen model validation, while avoiding overly strong claims if the additional analyses provide only modest support for a single preferred model.

  6. Version published to 10.64898/2026.03.04.709582 on bioRxiv