Confidence in Moral Decision-Making

Human beings are not only capable of making choices but also evaluating the quality of their choices. This evaluation process is related to a feeling of subjective confidence in choice (Allen et al., 2016; Sharot et al., 2023). Subjective measures of confidence have interested psychologists for centuries (Peirce & Jastrow, 1884). Behavioral scientists have measured the confidence of experts and laypeople alike, from the certainty of meteorologists in their weather forecasts (Murphy & Winkler, 1977), to the confidence of eyewitness identifications in criminal investigations (Wixted & Wells, 2017), to the assuredness of students in their test scores (Fischhoff & MacGregor, 1982). Within a decision-making context, confidence indexes the positive, subjective sense that an individual has regarding the quality, correctness, or accuracy of their judgment or choice (Rouault et al., 2023). In cognitive science, confidence has frequently been studied in the fields of perceptual judgment and value-based decision-making (Clairis & Pessiglione, 2022; Lebreton et al., 2015). Prior work across these domains has behaviorally and computationally dissociated first-order choices (e.g., identifying a stimulus as red as opposed to green, or selecting an apple over a banana) from second-order, introspective senses of confidence (Fleming et al., 2012; Folke et al., 2016; Sakaki et al., 2024).

Early work on the cognitive mechanisms underlying the formation of confidence has been focused on perceptual decision-making. For example, in a widely used perceptual decision-making task, participants are presented with an array of randomly moving dots on a screen, in which a proportion of dots (e.g., 10%) move coherently in the same direction (Fetsch et al., 2014; Gold & Shadlen, 2007). Participants not only decide which direction the coherent dots move (e.g., left vs. right), but also evaluate how confident they are for each of their decision (e.g., ref (Lehmann et al., 2022; Rollwage et al., 2020; Rouault et al., 2019)). A dominant computational model for behaviors in perceptual tasks, the drift-diffusion model, describes the cognitive processes underlying perceptual decision-making as a process where evidence in favor of each choice accumulates until a decision is made (Herz et al., 2016; Zylberberg et al., 2012). Within perceptual tasks, confidence has been connected to the evidence in favor of the chosen option – when the perceptual evidence for one option (e.g., more dots coherently moving towards left) is more obvious, the confidence for that choice is also higher (Zylberberg et al., 2012).

People make value-based decisions (e.g., between an apple and an orange) by computing and comparing subjective values associated with each available option (Chib et al., 2009; Glimcher, 2004; Hutcherson & Tusche, 2022; Rangel et al., 2008). Unlike perceptual choice tasks, value-based choice tasks typically contain perceptually unambiguous stimuli (De Martino et al., 2013; Folke et al., 2016). Consider a study in which participants are asked to choose between pairs of food options before subjectively rating their confidence. A participant who makes the decision between an apple and an orange should assume that these two choices are unmistakably and concretely defined. As such, confidence ratings should not reflect uncertainty about physical or perceptual distinctions between both fruits. Instead, in value-based decision-making, the main source of uncertainty involved in the decision process, and thus the primary factor influencing confidence, is the relative subjective values of the available options.

To compute confidence for value-based decisions, participants have to integrate across a number of component cognitive processes that reflect internal psychological states (e.g., hunger, preferences, beliefs) (Folke et al., 2016). As in the perceptual decision-making context, the more different the options are in terms of their subjective values (i.e., more evidence for one over the other options), the more confident participants are about their choices (Folke et al., 2016; Pisauro et al., 2017). For example, Folke and colleagues found that participants were more confident when the absolute difference in value between the available options (food items) was high and when behavioral reactions were faster (Folke et al., 2016). Notably, lower confidence ratings at the time of the initial choice were associated with subsequent changes of mind (i.e., participants later selected a different food item) when the same choice set was presented again (Fleming et al., 2018).

In the unified, “common currency” theoretical framework of value-based decision-making, social and moral decisions are also made based on subjective values of available options that have social and moral implications (e.g., harming an innocent person for monetary profits). In this framework, individuals making social and moral decisions are hypothesized to go through similar valuation processes as in non-moral value-based decision-making context (Lockwood et al., 2020; Ruff & Fehr, 2014). Most notably, in this framework, material (e.g., monetary profits) and non-material goods (e.g., being honest) are converted into an abstract, integrated value parameter and compared across all options (Crockett, 2016; Hutcherson et al., 2015; Van Bavel et al., 2015; Yu et al., 2019). It is, therefore, possible that confidence in moral decision-making follows the same principle underlying confidence in non-moral value-based decision-making, namely, confidence is positively associated with distance between the options in terms of subjective value.

To investigate the cognitive processes underlying confidence in moral decision-making, we adopted an established value-based moral decision-making task, which enabled us to manipulate, in a trial-by-trial manner, decision attributes and obtain measures of participants subjective values in moral decision-making (Crockett et al., 2014, 2017; Crockett & Cools, 2015; Yu et al., 2021, 2022). In the original version of the task, participants make a series of binary choices between a high-harm option, which contained more electric shocks and amounts of money, and a low-harm option, which contained fewer shocks and less money. The money was always for the participant (i.e., Decider), but the shocks were allocated to an anonymous Receiver in half of the trials (i.e., Other-condition) and to the participant in the other half of the trials (i.e., Self-condition). By independently manipulating the number of shocks and amount of money, and therefore sampling across the decision space, this task allows us to quantify participants’ relative harm aversion for themselves and another person. In Study 1, we adopted a hypothetical version of this task and asked the participants to report their confidence following each hypothetical decision. In Study 2, we adapted this task to include non-physical harm, to test the robustness and generalizability of our findings from Study 1. We also added a potential incentive for participants’ choices by providing a real monetary bonus (i.e., Amazon gift card, see below). Specifically, the participant (i.e., Decider) made a series of binary choices between a high-harm option, which contained a higher value Amazon gift card at a higher monetary cost, and a low-harm option, which contained a lower value Amazon gift card at a lower monetary cost. The Amazon gift card was always for the participant, but the monetary cost was sometime charged to the participants’ own initial endowment and sometimes to the Receiver’s endowment. Hence, the Amazon gift card was analogous to profits in the original task, and the monetary cost was analogous to the electric shocks.

We hypothesized that, based on work from non-moral value-based decision-making, participants’ confidence would be positively associated with the subjective value difference between the two options. Because decisions affecting others are inherently uncertain (Delton et al., 2011; Harsanyi, 1977; Rand et al., 2012; Siegel et al., 2018), we also hypothesized that participants would be overall less confident in the other-condition than in the self-condition. In addition to these hypothesis-driven analyses, we also explored how confidence varied as a function of underlying decision attributes (i.e., harm and profits) and how participants’ choices (high-harm vs. low-harm) modulated the formation of confidence. We are interested in understanding the effects of choice on participants confidence because previous studies using this task have consistently demonstrated that choosing the high-harm option, relative to the low-harm option, elicits blameworthy judgments from observers as well as the Deciders themselves (Crockett et al., 2017; Siegel et al., 2017; Yu et al., 2022). It is thus conceivable that Deciders are aware that choosing the high-harm option is generally more blameworthy than choosing the low-harm option. It is still an empirical question whether and how such implicit moral evaluation of the choice influences individuals’ confidence in their moral decision-making.

The study material, data, and analysis codes used in the present study are available on this OSF account: https://osf.io/4znr2/.

Two hundred and forty-eight U.S. participants were recruited on Amazon Mechanical Turk (mTurk) in April 2019. Participants were compensated at a base rate of $7.25 per hour. Participants who failed the attention checks (N = 6) or exclusively selected either the right or left option for the duration of the task (N = 2) were excluded from the final analysis, leaving a final sample of N = 240 (mean age: 31.8 years; age range: 18 – 74 years; 156 male, 81 female, 3 non-binary; 168 White, 31 Black, 26 Hispanic, 7 Asian, 8 others). The study was not pre-registered, and no prior power analysis was conducted. We used simple correlation as an approximation of the regression analysis for effect size estimation and conducted a sensitivity analysis. Based on the main hypothesis of a positive linear relationship between subjective confidence and the absolute difference in subjective value, the sensitivity analysis revealed that the smallest detectable effect size with power > 0.9 is 0.25. This is smaller than the observed effect size of 0.28. Therefore, Study 1 demonstrated sufficient power to detect the smallest effect size.

Each participant completed a Qualtrics survey. Before completing the primary decision task, participants provided their informed consent. The procedure was approved by the Human Subject Committee of Yale University (HSC#: 2000022385). All methods were carried out in accordance with the approved protocol.

Participants then completed the primary decision task, which was a hypothetical adaptation of an established laboratory-based task (Crockett et al., 2014). In the present study, participants were asked to imagine themselves in a laboratory setting. Here, participants were always assigned the role of the Decider and were told to imagine that their partner was another anonymous participant. As in the original study, on each trial, participants made a choice between a relative high-harm option and a relative low-harm option (Figure 1 ). The high-harm option contained more money and more electric shocks than the low-harm option. As in previous studies, we manipulated the amount of money and number of electric shocks such that the difference in monetary gain between the two options (i.e., ∆g, where g stands for “gain”) and the difference in the number of shocks between the two options (i.e., ∆l, where l stands for “loss”) were uncorrelated across the trials. Additionally, no trials included an option with a value of 0. This means that participants were not presented with an easy choice. Three trial sets were created and participants were randomly assigned to one of them (cf. Crockett et al., 2014; Contreras-Huerta et al., 2022). The detailed procedure of trial sets determination was described elsewhere (Crockett et al., 2014; Yu et al., 2021). The trial sets can be found on the OSF page associated with this paper, and distributions of ∆g and ∆l are provided in the Supplemental Materials (Figure S1, Figure S2) along with their descriptive statistics (Table S1). The rationale of including three different trial sets was to ensure that results are generalizable. Moreover, the high-harm and low-harm options were randomly placed on either the left or right. Participants made a total of 70 choices, half of which were self-trials (shocks for the Decider) and half of which were other-trials (shocks for the Receiver). As in the original experiment, the money always went to the Decider. After each decision, participants would indicate their subjective confidence in the decision they just made (“How confident are you in your choice?”) using a sliding scale (0 = “Not At All Confident”, 100 = “Extremely Confident), which appeared on a separate page of the survey.

In the present study, participants were asked to make each decision as if it were real. Participants were provided with the same context as in the original laboratory-based experiments; the shocks were described as “unpleasant, but not unbearable” and “comparable to a painful shot at the doctor’s office or a sharp, prolonged pinch.” It was made explicit that the shocks do not cause lasting harm or damage. The full text of the instructions that the participants read before the experiment can be found in the Supplemental Materials.

Overall, the participants chose the high-harm option more in the self-condition than in the other-condition (B±S.E = 0.20±0.03, z = 6.02, p < 0.001; for details, see Supplemental Materials: Analysis of Participants’ Actual Decisions). We modeled participants’ choices using an established utility model for this moral decision task (Crockett et al., 2014, 2017; Crockett & Cools, 2015) (Eq. 1). In essence, this utility model quantifies the exchange rate between money and pain for each participant separately for the self and the other-condition. Specifically, the subjective value of choosing the high-harm option (over the low-harm option) is a linear function of the relative money difference and shock difference on a given trial,

Here Δg and Δl represent the money (i.e., gain) difference and shock (i.e., loss) difference associated with the high-harm relative to the low-harm option. The harm aversion parameter, κ, reflects the exchange rate between shocks and money. Previous studies using this task have consistently shown that participants’ decision preferences are different when harming themselves for profit versus harming another person. Therefore, we allowed κ to vary across the self and other conditions (κ_self and κ_other). To fit participants’ binary choice data, we used a softmax function to translate ΔV into a probability of choosing the high-harm option over the low-harm option,

Here, β is a participant-specific inverse temperature parameter that characterizes the steepness of the softmax curve, or conversely, the noisiness of the participants’ choices.

The model correctly predicted 77% of participants’ choices (confidence interval of accuracy [73%-80%]), comparable to previous studies using this model (Crockett et al., 2014, 2017; Yu et al., 2021). With this modeling procedure, we replicated key findings from past studies using this task (Crockett et al., 2014, 2017; Volz et al., 2017; Yu et al., 2022) - on average, participants were more averse to harming others than harming themselves for money (𝜅_self: M±S.E. = 0.41±0.02; 𝜅_other: 0.46±0.02; t₂₃₉ = -2.06, p = 0.041).

We estimated a linear mixed effect model where we included trial-by-trial confidence ratings as the dependent variable. We included the main effects of absolute difference in subjective value between the choice options (|ΔV|), condition (self vs. other), and choice (high vs. low harm), and all two-way and three-way interactions as fixed effect predictors. The participant was included as a random intercept. For all the mixed effects regression models we reported here and below, we also estimated a different version where all the predictors were included as random slopes. Adding these random slopes only changed the significance of the condition (self) predictor for Study 2 in the model where confidence was predicted by subjective value difference. Since these models had converging issues, we reported the results of these models in the Supplemental Materials (Table S2, Table S3).

Since participants’ choices (high-harm vs. low-harm) varied as a function of condition, they were not completely independent. To estimate the degree of multicollinearity, we calculated the Variance Inflation Factor (VIF) for each predictor in the model. The main effect predictors (condition and choice) had VIF values below 5, indicating minimal collinearity. The VIF value of the choice-by-condition interaction was 5.74, indicating minimal to moderate collinearity (for VIFs of all the models we reported in this paper, see Table S6 and Table S7). However, since the interaction term tests how condition moderates or modulates the effect of choice on confidence, we believe that the level of multicollinearity observed in here and subsequent models does not significantly distort our statistical inference. For example, statistical work has demonstrated that multicollinearity of moderation (or interaction) terms does not have any meaningful impact on moderated regression models (McClelland et al., 2016).

First, we examined the relationship between subjective value difference and subjective confidence. As predicted, confidence was positively associated with |ΔV| (B±S.E. = 0.30±0.04, 95% CI = [0.23, 0.38], b = 0.08, t = 7.92, p < 0.001), indicating that when the subjective value difference between the two options was high, participants reported higher choice confidence (Figure 2 ). Next, we examined the effect of the harm recipient on confidence. This effect was significant (B±S.E. = 1.28±0.36, 95% CI = [0.57, 2.00], b = 0.04, t = 3.52, p < 0.001), indicating that participants were overall less confident when their choices affected another person than when their choices affected only themselves (Figure 2 ). No other predictors had a significant effect (Table 1 ).

We conducted some exploratory analyses to investigate how confidence varies as a function of underlying decision attributes. We ran a model where trial-by-trial confidence ratings were included as the dependent variable. Absolute difference in monetary gain (Δg), absolute difference in loss (Δl), condition (self vs. other), the choice on a given trial (high vs. low harm), and all the two-way and three-way interactions were included as fixed effect predictors.

The three-way interaction of Δl, condition, and choice was significant (B±S.E. = 0.188±0.062, 95% CI = [0.065, 0.311], b = 0.07, t = 3.011, p = 0.002) (Figure 3A ). By definition, when the high-harm option was chosen, the participants would make less relative harm as the absolute difference in shocks decreased. At the same time, participants expressed more confidence, both when the harm was hypothetically delivered to the Receiver (B±S.E. = -0.143±0.030, 95% CI = [-0.202, -0.084], b = -0.05, t = -4.74, p < 0.001) and to themselves (B±S.E. = -0.294±0.029, 95% CI = [-0.350, -0.237], b = -0.10, t = -10.16, p < 0.001). Critically, however, the increase in confidence was smaller when the harm was hypothetically delivered to the Receiver (i.e., immoral) than when the harm was hypothetically delivered to themselves (i.e., amoral), as indicated by a significant interaction between the recipient of harm and the absolute difference in loss (B±S.E. = -0.164±0.042, 95% CI = [-0.246, -0.081], b = -0.06, t = -3.88, p < 0.001) (Table 2 ). This interaction pattern was not observed when the participants chose the low harm option (B±S.E. = 0.035±0.046, 95% CI = [-0.054, 0.124], b = 0.01, t = 0.768, p = 0.443)

The three-way interaction of Δg, condition, and choice was not significant (B±S.E. = -0.02±0.06, 95% CI = [-0.15, 0.11], b = -0.007, t = -0.301, p = 0.763). However, the two-way interactions between Δg and choice (B±S.E. = -0.44±0.047, 95% CI = [-0.53, -0.35], b = -0.15, t = -9.442, p < 0.001), and between Δg and recipient (B±S.E. = 0.094±0.042, 95% CI = [0.01, 0.17], b = 0.03, t = 2.205, p = 0.027) were significant (Table 2 ). As can be seen from Figure 3B , when the participants chose the high-harm option, the more additional money they obtained from their choice, the more confident they were. In contrast, when the participants chose the low-harm option, the more additional money they could have obtained, the less confident they were. Furthermore, when the recipient of harm was themselves, their confidence was higher across all choices.

Participants in Study 1 made hypothetical moral choices in a physical harm context; critically, not all moral decisions involve physically harming others. Importantly, research in the cognitive science of morality has found that spatial proximity, personal force, and physical contact have differential effects on moral acceptability judgments (Greene et al., 2009). As such, it may be that moral choices which vary in terms of physical and non-physical impact also differ in how confidence about the choices is constructed.

To address the question of how moral confidence varies across physical harm and non-physical harm contexts, we conducted a follow-up study that was similar in task structure but different in terms of this particular aspect of moral consequence. Here, we conceptualized taking money from another person as a form of non-physical harm, since this act can be construed as theft. Participants made incentivized choices about how to spend their own or another person’s $20 endowment in exchange for an Amazon gift card for the participants. Furthermore, participants were informed that one pair of participants would be randomly selected and one of the Decider’s choices would be randomly selected and made real (see Supplemental Materials for the exact wording of the instructions).

U.S. participants were recruited on Prolific (https://www.prolific.co/) in August 2020 and were compensated at a base rate of $7.25 per hour. Participants who failed two attention checks (N = 38) were excluded, resulting in a final sample size of N = 369 (mean age: 27.8 years; age range: 18 – 70; 185 male, 176 female, 6 non-binary, 2 no answer; 277 White, 16 Black, 36 Hispanic, 25 Asian, 15 others). As in Study 1, Study 2 was not pre-registered, and no prior power analysis was conducted. However, a sensitivity analysis revealed that the smallest detectable effect size with power > 0.9 is 0.188, which is comparable to the observed effect size of 0.180. Therefore, Study 2 had sufficient power to detect the smallest effect size.

The experiment was implemented in Qualtrics. Before completing the primary decision task, participants provided their informed consent. The procedure was approved by the Human Subject Committee of Yale University (HSC#: 2000022385). All methods were carried out in accordance with the approved protocol.

Participants then completed the primary decision task, which was an adaptation of Crockett et al. (2014) in a monetary decision-making context. In the present study, participants were asked to imagine themselves in an experimental setting, where both they and another person were each given a $20 endowment. As in Study 1, participants were always assigned the role of the Decider and were told to imagine that their partner was another anonymous participant. Participants then made a series of decisions between giving up a small amount of endowment for a small Amazon gift card or giving up a larger amount of money for a larger Amazon gift card.

Sometimes, the participant (‘Decider’) gave up some amount of their $20 endowment in exchange for the gift card (i.e., self-trials), and sometimes the payment came out of the other person’s (‘Receiver’) $20 endowment (i.e., other-trials) (Figure 4 ). The trial sets were generated in a similar procedure as in Study 1. Analogously, the amount of money loss was treated as a “monetary harm”, akin to the physical harm (i.e., electric shocks) in Study 1, and the amount of gift card obtained was treated as monetary gain. To be consistent with Study 1, we referred to the option with relatively higher monetary losses on each trial as the high-harm option, and the other as the low-harm option. Similarly, the difference in the monetary loss between the two options was labeled Δl and the difference in gift card between the two options was labeled Δg. The survey was generated such that the high-harm and low-harm options were randomly placed on either the left or right side of the screen. Participants made a total of 80 choices, half of which were self-trials and half of which were other-trials. As in Study 1, the Amazon gift card always went to the Decider. The trial sets adopted in Study 2 can be accessed on the OSF page of this paper. The distributions of Δg and Δl, along with their descriptive statistics, are provided in the Supplemental Materials.

Participants also used a sliding scale – which ranged from 0 (“Not At All Confident”) to 100 (“Extremely Confident”) – to subjectively rate their confidence after each trial (Figure 4 ). The choice component and confidence component of each trial were exhibited on separate pages of the survey.

We applied the same computational model to participants’ choices as we did in Study 1. In this context, higher 𝜅 indicates that the participants were willing to give up less amount of the endowment in exchange for the Amazon gift card (i.e., avoiding monetary harm). The model correctly predicted 88% of participants’ choices (confidence interval of accuracy [87%-90%]), comparable to previous studies using this model (Crockett et al., 2014, 2017; Yu et al., 2021). Surprisingly, we found that the average harm aversion for self (𝜅_self, 0.514± 0.005) was higher than the average harm aversion for others (𝜅_other, 0.462± 0.010; t₃₆₈ = 5.16, p < 0.001), which was inconsistent with both Crockett et al. (2014) and Study 1 of the present paper. This indicates that people are not averse to causing monetary loss to others in the same way as they are averse to causing physical harm to others.

To examine the relationship between confidence and absolute difference in subjective value, we estimated a linear mixed-effect model where we included trial-by-trial confidence ratings as the dependent variable (Figure 5 ). We included as fixed effect predictors the main effects of absolute difference in subjective value (|ΔV|), condition (self vs. other), and choice (high vs. low harm), and all two-way and three-way interactions. The participant was included as a random intercept.

Replicating previous findings using value-based decision-making tasks and Study 1, confidence was positively associated with |ΔV| (B±S.E. = 0.90±0.05, 95% CI = [0.80, 1.003], b = 0.12, t = 17.11, p < 0.001). The main effect of condition (i.e., harm recipient) remained significant in this model (B±S.E. = 1.25±0.42, 95% CI = [0.43, 2.07], b = 0.13, t = 3.005, p = 0.003). Interestingly, the three-way interaction was significant (B±S.E. = 1.01±0.15, 95% CI = [0.71, 1.31], b = 0.14, t = 6.64, p < 0.001) (Table 1: Study 2). To unpack this three-way interaction, we examined the interaction between absolute difference in subjective value and choice separately for the self-condition and the other-condition. In both conditions, the positive association between confidence and absolute difference in subjective value (|ΔV|) was weaker when a high-harm (relative to low-harm) option was selected, as indicated by the significant interaction between |ΔV| and choice (Self: B±S.E. = 1.42 ±0.12, 95% CI = [1.17, 1.67], b = 0.15, t = 11.27, p < 0.001; Other: B±S.E. = 0.61±0.91, 95% CI = [0.42, 0.78], b = 0.10, t = 6.61, p < 0.001). This suggests that the translation from |ΔV| to confidence is less “efficient” (i.e., flatter slope) when a high-harm option is selected. Importantly, this difference was significantly larger when the recipient of the (monetary) harm was the Receiver than when the recipient was the participants themselves, as indicated by the significant three-way interaction. This suggests that causing harm to another person (i.e., an immoral action) may suppress the cognitive process through which the subjective value difference is translated into confidence. An alternative possibility is that the participants were more conflicted when they made decisions in the other-condition (e.g., making trade-off between obtaining material profits and violating moral rules) (Yu et al., 2022), and this conflicted cognitive process interferes with meta-cognition and confidence (Xu et al., 2024).

As in Study 1, we estimated another linear mixed effect model, where we examined how confidence varied as a function of the underlying decision attributes - Δl and Δg (Figure 6 ).

As in Study 1, we found a significant three-way interaction of Δl, condition, and choice (B±S.E. = 1.14±0.094, 95% CI = [0.96, 1.33], b = 0.29, t = 12.21, p < 0.001) (Figure 6A). When the high-harm option was chosen, the participants would make more monetary harm as the absolute difference in monetary cost increased, and they became less confident about their choice (main effect of Δl in the high-harm trials: B±S.E. = -0.374±0.034, 95% CI = [-0.44, -0.31], b = -0.09, t = -10.88, p < 0.001). Critically, such a decreasing slope was significantly steeper when the recipient of harm was the Receiver than when the recipient was the participants themselves (Δl-by-recipient interaction in the high-harm trials: B±S.E. = 0.187±0.056, 95% CI = [0.08, 0.30], b = 0.05, t = 3.36, p = 0.008), suggesting the modulation of moral status of choice (immoral vs. amoral) on the appraisal of confidence. When the low-harm option was chosen, the participants were more confident as the degree of harm they avoided increased (B±S.E. = 0.376±0.038, 95% CI = [0.30, 0.45], b = 0.09, t = 10.01, p < 0.001). Interestingly, the increase in confidence was smaller in extent when the participants avoided more monetary harm to the Receiver (i.e., a morally praiseworthy choice) than when they avoided the same degree of monetary harm to themselves (i.e., a morally neutral choice), as indicated by the significant Δl-by-recipient interaction (B±S.E. = 0.792±0.050, 95% CI = [0.69, 0.89], b = 0.20, t = 15.74, p < 0.001).

The model also revealed a significant three-way interaction of Δg, condition, and choice (B±S.E. = -0.645±0.086, 95% CI = [-0.81, -0.47], b = -0.19, t = -7.49, p < 0.001) (Table 2: Study 2). To unpack this three-way interaction, we estimated two regression models separately for high-harm choices and low-harm choices. In these models, trial-by-trial confidence rating was included as the dependent variable, harm recipient (self vs. other), Δg, and their interaction were included as fixed effect predictors. Again, participant was included as random intercept. For both high-harm trials and low-harm trials, the interaction terms were significant (Low-harm: B±S.E. = 0.511±0.056, 95% CI = [0.40, 0.62], b = 0.12, t = 9.085, p < 0.001; High-harm: B±S.E. = 0.294±0.044, 95% CI = [0.21, 0.38], b = 0.09, t = 6.63, p < 0.001). However, by inspecting their confidence intervals, the interaction effect was significantly larger in the low-harm trials than in the high-harm trials. From Figure 6B , it is also clear that the pattern of interaction is different in these two types of trials. In the high-harm trials, both regression lines are positive, indicating that the more monetary profit the participants obtained (by choosing the high-harm option), the more confident they were about their choice. But this is only true when the participants themselves were the recipients of the (monetary) harm. When their high-harm choices had moral consequences (i.e., in the other-condition), the participants confidence did not go up significantly as they obtained more monetary profit. For low-harm trials, although the direction of the interaction was the same as in the high-harm trials, the pattern suggests a different meaning. Here, the more monetary profit the participants gave up (as Δg increases), the less confident they were overall. But this was only the case when the harm recipient was the Receiver.

Across two hypothetical moral decision-making tasks, we demonstrated that participants reported greater confidence when the two options differed greatly in terms of subjective value, replicating the findings from studies of non-moral value-based decision-making (De Martino et al., 2013; Folke et al., 2016; Lebreton et al., 2015). This effect was consistent across both studies, highlighting the robust relationship between subjective value differences and confidence. Moreover, in line with previous social psychological research and theories, making decisions with consequences for oneself is associated with more confidence than making decisions that affect others (Delton et al., 2011; Harsanyi, 1977; Siegel et al., 2018). This finding was consistent across both studies, emphasizing that self-related decisions inherently carry more confidence.

Importantly, we also found evidence for unique features of the computation of confidence in the moral decision-making context: the moral status of choice modulates the conversion from decision attributes to confidence. A morally blameworthy act – causing harm to an innocent person (relative to oneself) for one’s own profits – suppressed the conversion of gaining more profits to more confidence, while making the degree of harm a stronger dampening factor of confidence. This nuanced effect was particularly evident in Study 2, where several interaction terms involving choice and condition significantly impacted confidence.

Besides the commonalities across the two studies, some differences were also observed. The choice (high vs. low-harm) showed significant modulation effects (i.e., two-way and three-way interactions) only in Study 2. This indicates that participants are generally less confident when choosing a high-harm option compared to a low-harm option (i.e., the main effect of choice). However, this reduction in confidence from high-harm choice to low-harm choice is even more pronounced when the recipient of the harmful consequence is another person than themselves (i.e., the two-way interaction between choice and condition). Moreover, subjective confidence is less sensitive to absolute value differences when choosing a high-harm option than when choosing a low-harm option. In other words, the conversion from absolute value difference to subjective confidence is less efficient when a high-harm option is chosen. This reduction in sensitivity is even stronger when the recipient of the harmful consequence is another person (i.e., the three-way interaction). Our previous work has shown that in a harm aversion task as the one we used here, choosing the high-harm option is seen as more morally blameworthy when the recipient of the harmful consequence is another person than when the recipient of the harmful consequence is the decider themselves (Crockett et al., 2017; Yu et al., 2022). These findings suggest that while certain factors, such as subjective value differences and recipient of the consequences, consistently impact confidence, the moral context and specific framing of choices introduce additional layers of complexity. The observed differences between the two studies may be attributable to multiple factors (e.g., the differences in the incentive nature and type of harm involved in the two studies) and highlight the need for further research to uncover the underlying mechanisms driving these variations. Understanding why certain effects are more pronounced in one study compared to the other could provide deeper insights into the cognitive and contextual factors that shape moral decision-making processes.

Using the same moral decision-making tasks, Crockett and colleagues showed that the same amount of monetary gain produced weaker neural responses in the brain valuation system when it was obtained through harming another (i.e., morally blameworthy) than harming oneself (i.e., morally neutral) (Crockett et al., 2017). This result provides a neurocognitive account of why people avoid harming others even more than avoiding harming themselves – the profits obtained through immoral means are less valuable. Expanding on this finding, we showed that profits obtained through immoral means also produced less subjective confidence. Together with previous work showing that lower confidence is a predictor of a “change of mind” in the future (Fleming et al., 2018; van Den Berg et al., 2016), our results suggest that low confidence in past immoral behaviors may give rise to morally neutral or good behaviors in the future. Future research with designs that could better capture inter-trial dependency is needed to rigorously test this possibility.

There are several limitations in the present studies. Both studies were conducted online and asked participants to imagine themselves in a decision-making situation. While Study 2 involved the possibility of real monetary incentives, the study environment is different from a laboratory setting where the Decider participants have some interactions with the Receiver participants as in some previous studies using the moral-decision making task (e.g., during random assignment) (Crockett et al., 2014; Siegel et al., 2018; Volz et al., 2017; Yu et al., 2021, 2022). This limitation stems from both tasks requiring participants to make decisions that would negatively affect another ‘participant’, who was neither physically present nor actually real in either study. It is conceivable that participants would be less likely to harm another participant if this study was conducted in a laboratory setting. Indeed, previous work in moral decision-making has found that real and hypothetical tasks not only trigger distinct patterns of behavior, but also engage different neural structures (FeldmanHall et al., 2012; Yu et al., 2024). Even if participants’ decisions remained unchanged, their confidence in these choices could be altered if this experiment were conducted in person and had more emotionally tangible or salient consequences.

The two experimental tasks differ in how easily comparable the options are, or how ‘computable’ subjective value comparisons are between choices. It is highly unlikely that, before this task, any participants have had to make decisions about how to physically harm themselves or another person in exchange for money. This computation, therefore, is novel, and participants are less likely to come in with preexisting beliefs about what a ‘proper’ exchange rate for such a decision would be. Further, while both the money and shocks are presented with precise quantitative values, it is inherently challenging to put a price on pain – herein lies the precise logic of this paradigm. In contrast, the task structure of Study 2 lends itself to more familiar and day-to-day calculations. Most participants are likely familiar with the idea of an exchange rate and understand how credit can be stored on gift cards. Indeed, there exist entire marketplaces dedicated to buying and selling unwanted gift cards for discounted prices (GiftDeals.com, accessed August 7, 2023). Participants still need to make calculations about how they value untethered money relative to Amazon gift cards, and how this relationship changes when they are taking money from another person. Nevertheless, while the tasks intentionally differ in their harm-related consequences, they also likely differ in the ease with which participants can actively represent subjective values. As such, differences in Study 1 and Study 2 could potentially represent differences in their ease of computability rather than their relative harm contexts. Future studies that equalize the mental abstraction is needed to demonstrate the robustness of our findings.

If these subjective value comparisons are indeed easier to make in Study 2 relative to Study 1, then the participants might be able to better strategically modify their choices (e.g., to avoid inflicting harm or to maximize personal profit) or modulate their reported confidence (e.g., for social desirability) as a reaction to their improved representations of their choices. Critically, a growing body of work has found connections between confidence and post-decision evidence accumulation and biases, which cements this concern as a realistic confound (Fleming & Daw, 2017; Navajas et al., 2016).

Relatedly, neither study incentivizes a precise or truthful disclosure of confidence. Although Study 2 adopted an incentivized task, the incentive was uncertain (i.e., only one participant would be randomly selected to receive the gift card) and not contingent upon the truthfulness of their self-reported confidence. Future research should identify a way to incentivize truthful and precise confidence reports within the moral decision-making domain. One possibility is to integrate participant’s confidence ratings into the process used to determine which singular trial is ultimately carried out. For example, perhaps some fixed number of random trials could be selected from the overall set. Then, participants’ confidence ratings of this subset could be used to probabilistically determine which trial is carried out. Such a process would incentivize participants to report their truest, knowable subjective confidence without allowing them to ‘hack’ the outcome. Relatedly, in the present research, we examined participants’ general, holistic sense of subjective confidence about their choice overall, rather than confidence in any specific aspects of their choice, such as confidence about whether the choice was morally right, or whether it was rational (i.e., maximizing subjective value). Future research is needed to closely investigate these more specific aspects of confidence. For example, in addition to self-report, future research could adopt behavioral measures of confidence (e.g., “change of mind” when given a second chance) to better distinguish different aspects of confidence (Fleming et al., 2018; Folke et al., 2016; Rollwage et al., 2020; Sharot et al., 2023).

Another limitation of this study, and an important opportunity for future research, relates to reaction times. Prior work has shown that confidence is inversely linked with reaction times (Kiani et al., 2014). Notably, decision time itself affects confidence ratings without affecting choice accuracy, suggesting that people may use their decision time to estimate their performance on a complex task (Kiani et al., 2014). As such, future studies should either control reaction time or, alternatively, measure reaction times in a laboratory environment to account for the theory that people respond faster when they are more certain. Notably, Crockett and colleagues found that hyperaltruistic participants took more time to make decisions with consequences for other people than themselves (Crockett et al., 2014). Moreover, with both choice and reaction time data, researchers could obtain more information about the moral decision-making process itself, with the help of computational models such as the drift-diffusion model (Yu et al., 2021). The model-based characterization of the often-hidden decision-making process will provide a mechanistic account of confidence formation in moral decision-making context.

Finally, future studies should explore how moral confidence is constructed in other tasks or contexts. The tasks used in our studies centered around the concept of harm. However, harm may not be the only domain of moral cognition, such as loyalty and purity (Graham et al., 2013). It is a theoretically significant question as to whether choice confidence in different moral domains follow similar rules as in the harm domain.

To conclude, using model-based moral decision-making tasks, we identified important similarities in how moral confidence is constructed in moral decision-making context comparison with other decision-making contexts. We show that confidence is generally higher when people make decisions that have consequences for themselves as opposed to for other people, and when the options obviously differ in terms of the subjective value. We also show that moral status of choice modulates how decision attributes is converted to confidence – morally blameworthy behavior strengthens the effect of confidence-reducing decision attribute (e.g., harm), while dampens the effect of confidence-boosting decision attribute (e.g., profit).

S.H., M.G., H.Y., and M.C. designed the experiments, S.H. and M.G implemented the experiments, L.S., S.H., M.G., Y.Z., M.O., and H.Y. analyzed the data, L.S., S.H., and H.Y. wrote the initial draft, H.Y., M.O., and M.C. revised the manuscript.

This work was supported by grants from the John Templeton Foundation (Beacons Project and the Moral Narratives grant #61495) awarded to M.J.C. H.Y. was supported by a Theresa Seessel Endowed Fellowship, Yale University.

The author(s) declare no competing interests.

The data and analysis codes used in the present study have been made public on this OSF account: https://osf.io/4znr2/.