Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Communications Psychology volume 3, Article number: 143 (2025)
3503
21
Metrics details
How do people learn about their own abilities? Often, people receive rewards that offer information about their performance level. Yet, even when two people perform equivalently on a task, they may receive disparate rewards. In these cases, could rewards still influence self-evaluations of ability? In two behavioral experiments, we asked whether people feel more capable and confident when they receive more rewards, even when their performance is held constant, and they know how they objectively performed. Participants played a perceptual game in which they received trial-by-trial accuracy feedback; a staircase procedure held their objective performance constant. However, participants were assigned to either a high or low-reward condition, which varied the probability of receiving a reward for a correct answer. In Experiment 1 (N = 340), we found evidence that rewards bias overall self-evaluations of ability after the task—particularly estimations of objective accuracy. Next, in Experiment 2 (N = 342), we examined whether reward feedback would inflate participants’ trial-by-trial expectations of their own accuracy before each round of the game. Results indicated that participants updated their expectations to a greater extent when a correct response was accompanied with a reward. These findings suggest that rewards enhance how much people integrate accuracy feedback into their dynamic self-beliefs.
When learning new skills, people track their own performance, forming an impression of their own abilities. Once developed, these self-beliefs can influence the careers people pursue and the challenges they take on, shaping future performance, outcomes, and earnings1,2. One way people learn how they are doing on a given task is through positive feedback, such as good grades or performance bonuses, which can offer information about one’s performance level. However, the same level of performance may also garner different rewards. For instance, equally talented children may receive different amounts of praise, and equally skilled employees may receive different salaries. In these cases, rewards may still have an affective impact, leading people to feel more positively about their successes. By extension, would such variations in reward lead to biased perceptions of one’s own ability?
Beliefs about one’s own ability reflect one component of the self-concept, which encapsulates a broad collection of beliefs and attitudes about the self in various domains. The self-concept includes both stable components and temporary components that can be malleable in response to environmental influences3,4. As people form impressions of their ability in a task domain, these beliefs may serve as a momentary and temporary aspect of self-concept and, over many instances of learning, may eventually become integrated into more stable aspects of the self5.
Past work suggests that rewards can indirectly influence these self-beliefs by shaping behavior. For instance, greater financial incentives can motivate people to work harder and perform better in the workplace6,7. This behavior change can fuel a positive feedback loop: By motivating employees to work harder, rewards may lead them to develop an identity as a hard worker8. Yet, beyond reinforcing behavior, receiving rewards typically feels good. Affective experiences may, in turn, drive belief updating9,10,11,12,13,14. In particular, people are often unaware of the source of their feelings15, leading feelings to seep into judgments. For instance, individuals in a good mood tend to overestimate their overall life satisfaction16, and individuals who find instructions for a task easy to read tend to underestimate the difficulty of completing the task17. People may similarly attribute the positive feelings induced by rewards to their own ability when performing a task, leading them to form an inflated sense of their own performance. Indeed, in previous work, monetary incentives for accurate confidence judgments biased those judgments, such that the prospect of gains increased confidence and the prospect of losses decreased confidence18. This work underscores that rewards can elicit positive feelings that bias confidence in task-related performance estimates. Local decision confidence then shapes the formation of general self-beliefs about ability in a given task19. Yet, in addition to shaping immediate confidence in a particular response, rewards during learning about one’s actual performance in a task may lead to biased self-evaluations of one’s performance on the whole. Altogether, beyond indirectly influencing identity by reinforcing behavior, rewards may also directly influence self-beliefs through changing affect.
When forming impressions of others’ ability, rewards indeed influence evaluations of others’ competence. In one study, participants learned about others who varied both a) in their competence in answering trivia questions and b) in the amount of rewards they earned for participants through the game, which was determined randomly by lottery scratch-off tickets. Participants subsequently rated individuals who had earned them larger material rewards as more competent, even when these individuals had identical trivia performance to those who earned them lower rewards20. A similar mechanism may govern self-perception, whereby rewards might bias a person’s impression of their own competence even when they are aware of their objective performance.
Prior neuroimaging work similarly raises the hypothesis that rewards could become intertwined with expectations of success. Brain regions involved in reward processing often overlap with those that process accuracy feedback21,22, track perceptions of one’s own ability23, and form confidence judgments24,25. Thus, reward feedback may become conflated with ability as part of an expectation of success as people evaluate their task performance—a hypothesis we tested behaviorally in the present work.
Altogether, beyond indirectly shaping self-beliefs by reinforcing specific behaviors, reward may also directly mold self-beliefs by leading people to feel more positively about their performance. Although reward feedback can sometimes hold informative value about performance, we asked whether it has an influence even when a more objective metric of performance is readily available. In two preregistered experiments, we examined the impact of rewards on self-evaluations of ability by decoupling reward feedback from objective performance feedback. Using a between-subjects design, we manipulated the frequency of monetary rewards participants received for correct answers in a perceptual game. This game allowed us to hold objective performance (i.e., accuracy) constant and give explicit accuracy feedback after each round. In this way, participants could learn based on feedback about their performance in a task that did not clearly map onto pre-existing abilities, similar to prior paradigms used to study self-beliefs11,26,27,28. We measured the effect of rewards on participants’ (1) estimations of previous accuracy, (2) predictions of future accuracy, and (3) overall judgments of game competence. We refer to these measures as task-specific self-beliefs—measures reflecting impressions of performance, expectations of success, and generalized beliefs about ability in a task domain. Next, we examined the trial-by-trial mechanisms driving biased self-belief formation, testing whether rewards inflate immediate predictions of accuracy after the same accuracy feedback.
Given that participants received explicit feedback about accuracy, they could form self-beliefs by integrating across objective performance feedback. However, at times, people must assess their performance by reflecting on their experience—an ability known as metacognition. For instance, people may assess how confident they are in a response based on the experience of responding29. As a secondary question, we asked whether frequent rewards for correct answers would also shape later confidence in the absence of feedback. To do so, we included a separate phase of the game that measured trial-by-trial confidence in responses in the absence of external feedback. We refer to this measure as local confidence.
Across both studies, we hypothesized that participants who received more rewards during the task would estimate higher previous accuracy, predict higher future accuracy, and rate themselves higher in competence in the game. We also predicted that receiving more rewards would lead to having higher local confidence when completing the task again without feedback. On a trial-by-trial level (Study 2), we hypothesized that rewards would increase the influence of accuracy feedback on self-evaluations.
All manipulations and measures, sample size and exclusion criteria used in this study were preregistered on June 2, 2023. A preregistration document can be found at https://aspredicted.org/S2J_XSG. All data in Study 1 were collected in June 2023.
Participants. We employed a convenience sampling strategy and recruited 400 participants from CloudResearch with an approval rate of 95% or above. All participants were compensated $2.40 for their participation in the study, with bonuses beyond this base payment dependent on their performance in the game (these ranged from an additional 20 to 80 cents). 364 participants self-reported their age and gender. They ranged from 18 to 75 years (M = 39.86, SD = 11.46). 206 identified as men, 157 identified as women, and 1 as other. 360 reported their racial/ethnic affiliation (they could select multiple categories). The majority identified as White or Caucasian (68.6%), followed by Black or African American (10.8%), East Asian (8.2%), Hispanic or Latinx (6.4%), South Asian (2.3%), American Indian or Alaska Native (1.8%), Middle Eastern (0.8%), Other (0.8%), and Pacific Islander or Native Hawaiian (0.3%). The sample size was determined through a power analysis based on the exclusion rate and effect sizes observed in a pilot study (described in Supplementary Note 2). Participants were eliminated from analyses based on our preregistered rule: if they failed an attention check phase after the main task that included six rounds of the game at the easiest level (i.e., no noise in dot motion), defined as answering three or more rounds incorrectly. This left a total of 340 participants (85.0%) whose data were used in all Study 1 analyses (unless otherwise specified), with 176 participants in the high-reward condition and 164 in the low-reward condition. All participants completed informed consent in accordance with the University of Southern California Office of the Protection of Research Subjects, which approved the study procedures. We only included in our sample the participants who recorded task data; there were an additional 84 participants who started the Qualtrics survey (containing the study link) without participating in the experiment. In cases where multiple entries were recorded for the same participant, we only used their data from the first entry for analysis.
A post-hoc sensitivity analysis in G*Power 3.130 showed that the minimum detectable effect size at 80% power was f 2 = 0.018 for a single coefficient in a multiple linear regression model.
Procedure. We programmed the study using PsychoPy (version 2021.2.2)31 and set up a Qualtrics survey, which directed participants to the study link on the online platform Pavlovia. Thus, participants completed the experiment online without a researcher present. All participants played 80 rounds of the dot motion game, in which they saw dots moving on the screen and indicated the direction in which the majority of the dots were moving. On each round, the dots appeared for 0.75 seconds, after which subjects had unlimited time to indicate the direction of motion by pressing the left or right arrow keys. To ensure similar accuracy levels across all participants, we used a “staircase” procedure to adapt task difficulty and keep accuracy at roughly 70%32. This algorithm makes the task harder when a participant answers several rounds correctly and easier after incorrect rounds. Specifically, this meant adjusting the amount of random noise in the motion of dots, making it harder or easier to tell which direction the majority of dots were moving. We implemented a QUEST staircase using JsQuest33, which calculates the most probable Bayesian estimate of the difficulty level on each round that would yield the target accuracy rate of 70%34. Before starting the game, participants had 3 practice rounds, for which the task was at the easiest level (i.e., all dots moving in the same direction with no noise).
For every correct answer in the game, participants had a turn with a virtual slot machine; for incorrect answers, they got no turn with the slot machine. The virtual slot machine determined whether the participant received a token, and tokens translated to a monetary bonus on top of their base payment at a rate of 1 token = 1 cent. After each round in the game, participants received feedback both about their accuracy on that round as well as whether they earned a token in the case of a correct answer. This feedback appeared sequentially—first, accuracy feedback (i.e., “correct” or “incorrect” text) would display. In the event that participants had answered the round correctly, an image of a slot machine appeared underneath the accuracy feedback after 0.75 seconds. Then, reward feedback (i.e., an image of a token or no token) appeared 1.5 seconds after accuracy feedback, underneath the slot machine. All three items (accuracy feedback, slot machine, and reward feedback) would remain on screen for another second before the next trial, for a total of 2.5 seconds on this feedback page. If participants had answered the round incorrectly, only the accuracy feedback would display for the full 2.5 seconds.
Although answering correctly guaranteed participants a turn with the slot machine, it did not guarantee a reward. Participants were randomly assigned (via assignment logic within the experimental platform) to one of two between-subjects conditions. In the high-reward condition, the probability of the slot machine giving a token was 85%, while in the low-reward condition, the probability was 25%. We used these probabilities to generate strong and weak experiences of reward; specifically, a 70% accuracy rate from the staircase procedure would yield 56 trials on average with correct answers, and an 85% reward rate would yield approximately 48 rewarded trials, while a 25% rate would yield approximately 14 rewarded trials. In this way, reward delivery was probabilistic and unrelated to accuracy. After completing all rounds of the game, participants answered several questions about their previous performance. We used 3 measures to probe task-specific self-beliefs: estimated previous accuracy, predicted future accuracy, and self-evaluations of competence in the game. Participants rated their estimated previous accuracy and predicted future accuracy on continuous scales ranging from 0 to 100, corresponding to the percentage of rounds they believed they had answered correctly or would answer correctly in the future, respectively. For self-evaluations of competence, participants rated how good they were at the game on a continuous scale ranging from 0 to 100, with 0 as Not good at all and 100 as Very good. Additionally, participants also rated their enjoyment (i.e., “how much did you enjoy completing the dot motion game?”), perceived difficulty (i.e., “how difficult did you find the dot motion game?”), and motivation (i.e., “how motivated were you to do well at the dot motion game?”), all on continuous scales ranging from 0 to 100.
Local Confidence. We use the term “local confidence” in this study to refer to metacognitive judgments of confidence about one’s immediate performance at the trial level. In order to measure local confidence in immediate responses, participants played 20 additional rounds of the game. Participants received neither accuracy nor reward feedback during these trials. Trials were presented at the difficulty level at which the staircase had ended for each participant and were not subject to further staircasing, as prior work has suggested that using a single difficulty level provides the most accurate measure of metacognitive ability35. After each response, participants rated their confidence level on a scale ranging from 1 = Guessing to 4 = Very confident. Thus, these rounds provided a trial-by-trial measure of participants’ internal evaluations of confidence in a particular response while playing the game—as opposed to global evaluations of their performance—without influence from additional rewards or objective performance feedback. Past work has used this procedure for studying metacognitive judgments of confidence36. For this portion of the study, participants also learned that each correct answer earned one token; this instruction served both to incentivize performance as well as to establish a new reward structure in which previous reward contingencies became irrelevant. (See Supplementary Note 8 and Supplementary Fig. 2 for analyses of these data.)
Finally, participants played 6 additional rounds of the game at the easiest level. Like the previous trials, participants learned that each correct answer earned one token. We used performance on these trials to exclude inattentive participants. Since the easy trials occurred at the end of the experiment, we reasoned that performing at a chance level or below on these trials likely indicated a lack of learning or attention. We also had used a similar rule previously in a pilot study (reported in Supplementary Note 2). Further, to probe for knowledge about reward structure in the main phase of the game, participants estimated the probability with which they had previously received a token from the slot machine for a correct answer, forming a measure of perceived reward (see Supplementary Note 7 and Supplementary Table 11 for exploratory analyses associated with this measure). Since the 20 rounds probing local confidence and 6 additional rounds of the game did not have a slot machine that could give tokens for correct answers, we included these rounds before the questions probing perceived reward. However, if these additional rounds of the game had some biasing effect on estimates, this bias would have affected all participants regardless of the reward condition to which they were assigned. Finally, participants provided their demographic information and read a debriefing form regarding the study. See Fig. 1 for an overview of the experiment design.
Participants first practiced the game at the easiest level for 3 rounds before proceeding to 80 rounds of the main phase. Then, they filled out questionnaires probing task-specific self-beliefs. They then completed 20 additional rounds of the game without accuracy or reward feedback at the ending difficulty level from the main phase, rating their confidence after each answer. Finally, participants completed 6 rounds at the easiest level, which served as an attention check.
Analytic Strategy. We analyzed all data using R version 4.3.337. Graphs were generated using the ggplot238, ggdist39, and cowplot40 packages. We used multiple linear regression models to test whether reward condition (coded as −1 for low-reward and 1 for high-reward conditions) predicted estimations of previous accuracy, predictions of future accuracy, and competence ratings, controlling for objective accuracy. Even though the staircase procedure held accuracy at roughly 70%, we included objective accuracy as a covariate to account for any remaining variability in actual performance between participants and to confirm that participants had some veridical sense of how they had performed. For all multiple linear regression models reported, unless otherwise noted, residual plots showed no signs of major deviations from normality or evidence of heteroskedasticity, and variance inflation factors indicated no concerning multicollinearity, suggesting that model assumptions were met. In cases of deviation from normality or evidence of heteroskedasticity, we used the sandwich41 and lmtest42 packages to conduct regression with robust standard errors. For all t-tests reported, the data were visually inspected to be approximately normal. All statistical tests were two-tailed and used an alpha level of 0.05 to determine significance.
To test the effect of rewards on later local confidence, we fit a generalized estimating equation (GEE) model assuming a Poisson distribution (according to our preregistered plan) using the geepack package43. This approach assumes that the outcome is count-like, with equal spacing. Model diagnostics confirmed that the asymptotic and dispersion assumptions of the Poisson GEE were adequately met (dispersion ratio = 0.12). To ensure that the observed effect was robust to modeling assumptions, we also fit an exploratory cumulative link mixed model (including random intercepts for participants) using the ordinal package44, which is better statistically suited for modeling ordered categorical outcomes (results of this model are reported in Supplementary Note 8).
Experimental procedures, analyses, exclusion criteria for Study 2 were preregistered on September 8, 2023 on AsPredicted. A preregistration document and can be found at https://aspredicted.org/F92_NGB. All data in Study 2 were collected between September and October 2023.
Participants. We aimed to recruit 400 participants, but due to a platform error, data were collected from 399 unique participants. We employed a convenience sampling strategy. Of the recruited participants, 208 came from the University of Southern California subject pool and received course credit in exchange for participation in the study. 187 self-reported their age and gender. These respondents ranged from 18 to 44 years (M = 20.31, SD = 2.19), and 73 self-identified as men, 113 as women, and 1 as other. 184 participants also indicated their racial/ethnic affiliation, and they were allowed to select multiple categories. Of these participants, the most frequently selected categories were White or Caucasian (32.5%) and East Asian (31.1%), followed by Hispanic or Latinx (12.4%), South Asian (10.1%), Black or African American (4.8%), Middle Eastern (3.8%), American Indian or Alaska Native (1.4%), Pacific Islander or Native Hawaiian (1.0%), and Other (2.9%).
The remaining 191 participants came from CloudResearch and had an approval rate of 95% or above. These participants were compensated $2.40 for their participation in the study. 182 self-reported age, gender, and racial/ethnic affiliation; these respondents ranged from 19 to 72 years (M = 36.81, SD = 11.44), and 112 self-identified as men and 70 as women. The majority identified as White or Caucasian (60.1%), followed by Black or African American (16.6%), Hispanic or Latinx (8.8%), East Asian (8.3%), South Asian (3.6%), American Indian or Alaska Native (2.1%), and Middle Eastern (0.5%). No participants identified as Pacific Islander or Native Hawaiian or selected “Other”.
We used the same preregistered exclusion rule as Study 1 (excluding participants who answered incorrectly on three or more of the easy dot motion trials at the end of the experiment). After exclusions, a total of 342 participants (85.5%) remained for analysis, with 174 participants in the high-reward condition and 168 in the low-reward condition. All Study 2 analyses used this sample, unless otherwise specified. All participants completed informed consent in accordance with the University of Southern California Office of the Protection of Research Subjects, which approved the study procedures. We only included in our sample the participants who recorded task data; there were an additional 135 participants who started the Qualtrics survey without participating in the experiment. In cases of duplicate data entries for the same participant, we only used data from the first entry.
A sensitivity analysis indicated that the minimum detectable effect size at 80% power for a single coefficient in a multiple linear regression was f 2 = 0.018. The smallest detectable effect size for a two-tailed paired samples t-test was d = 0.16.
Procedure. As in Study 1, we programmed the study on Psychopy (version 2021.2.2)31 and created a Qualtrics survey that directed participants to the online study on Pavlovia. Participants completed the experiment online with no researcher present. The procedure in Study 2 was similar to Study 1, with the following changes. Before every round of the game, participants used the up or down arrow keys to indicate whether they expected that they would answer correctly or incorrectly on the next round, forming a measure of performance expectation. Additionally, because part of our sample in Study 2 came from a university subject pool, rewards did not translate directly to a bonus; instead, tokens for all participants served as raffle tickets for a $10 gift card draw at the end of the study. Moreover, in Study 2, each trial was associated with a unique image cue representing either a lottery ticket after a correct answer or an “incorrect receipt” after an incorrect answer. We used images of objects and scenes, taken from pre-existing image databases45,46 and randomized per participant whether objects represented lottery tickets and scenes represented incorrect receipts, or vice versa. We also randomized the order in which the images appeared. Following the game, we probed memory for several images by asking participants to indicate “Yes” or “No” in response to the question, “Did you see this image during the game?” 30 were new images that had not been displayed during the game, while we aimed to test up to 10 images previously associated with incorrect trials, up to 10 associated with correct-unrewarded trials, and up to 10 associated with correct-rewarded trials. We included these memory probes to test a second possible mechanism, whereby rewarded trials may stand out in memory, which could in turn serve to bias self-evaluations at the conclusion of the game. However, due to a critical data recording error, memory probe data could not be accurately associated with specific trials, preventing us from testing this alternative hypothesis. Before finishing the study, participants also filled out a general self-esteem questionnaire47 and the PHQ-9 to assess depressive symptoms48. (Analyses associated with the depression measure are reported in Supplementary Note 9). We added these measures to explore how individual differences may relate to task-specific self-belief updating.
Analytic Strategy. All analyses were implemented in R version 4.3.337, and graphs were generated using the ggplot238, ggdist39, and cowplot40 packages. We used a mixed-effects logistic regression using the lme4 package49 to test how trial-by-trial performance expectations (measured as a binary variable, where 0 = incorrect and 1 = correct; total of 80 trials per participant) depend on accuracy (i.e., whether participants answered correctly or incorrectly) and reward (i.e., whether reward was presented) feedback on the preceding trial. This model, therefore, estimated whether participants increased their expectations of success in response to correct (versus incorrect) answers and whether they did so to a larger extent if a correct answer was rewarded. The model included random intercepts for participant and by-participant random slopes for both reward and accuracy feedback. The correlations between random effects were not modeled. Model diagnostics indicated no violations of key assumptions, including linearity of the logit, multicollinearity, heteroscedasticity, or model convergence. The assumption of normality for t-tests was adequately met based on visual inspection of difference scores.
Identical procedures as Study 1 were used for modeling task-specific self-beliefs and trial-by-trial confidence. We confirmed that the assumptions of normality and absence of heteroskedasticity and multicollearity were met for multiple linear regression models. Model diagnostics for the GEE model predicting local confidence indicated that asymptotic and dispersion assumptions were met (dispersion ratio = 0.10).
We first investigated how the probability of receiving rewards during a task influences subsequent self-evaluations of ability. In a dot motion task described as a perceptual game, participants repeatedly judged whether dots were moving to the left or right. To hold performance approximately constant across participants ( ~70% accuracy), we implemented a staircase algorithm that increased or decreased the difficulty of the task, which was accomplished by increasing or decreasing random noise that made the direction of overall motion more or less difficult to detect34. After each round, participants received accuracy feedback, allowing them to track their performance objectively. In addition, participants were informed that each time they answered correctly, they would have the opportunity to play with a slot machine that could give them a token worth a monetary reward. On these rounds, they saw an image of a slot machine appear on screen, followed by a token or no token (Fig. 2). Thus, reward was probabilistic, and participants were randomly assigned to a “high-reward” condition, in which they received a token on 85% of correct trials, or a “low-reward” condition, in which they received a token on 25% of correct trials. Participants were instructed that, while correct responses always generated an opportunity to play with the slot machine, the outcome was unrelated to their response. After the game, they then made three self-evaluations about their own ability in the game: an estimate of the proportion of trials they had answered correctly, a prediction about the proportion of trials they would answer correctly if they played 100 additional rounds of the game, and a rating of how good (i.e., competent) they were at the game overall. We hypothesized that participants who received more rewards when completing a task would estimate higher previous accuracy, predict higher future accuracy, and rate themselves higher in competence in the game.
After making a response for a round of the dot motion game, accuracy feedback was displayed first, reading either “Correct” or “Incorrect.” If the response was incorrect, nothing else would display before the next round began. If the response was correct, an image of a slot machine appeared underneath the accuracy feedback, followed by an image of a token or a circle with a slash, indicating whether a token was won or not.
Although not specified in our preregistered plan, we examined whether the staircase procedure successfully held accuracy constant between reward conditions. The average accuracy was 69.32% (high reward: M = 0.69, SD = 0.09; low reward: M = 0.70, SD = 0.09), and a two-tailed independent samples t-test revealed no significant differences between reward conditions (t(337.8) = −1.86, p = 0.06, Cohen’s d = 0.20, 95% CI [−0.036, 0.001]).
We first assessed whether self-beliefs about the task differed across reward conditions, adjusting for each participant’s actual accuracy. The multiple linear regression models explained 11.1% of the variance in estimations of previous accuracy (Adjusted R2 = 0.111, F(2, 337) = 22.24, f 2 = 0.13, p < 0.001) and 13.9% of the variance for predicted future accuracy (Adjusted R2 = 0.139, F(2, 337) = 28.37, f 2 = 0.17, p < 0.001). Actual accuracy was significantly associated with both estimated previous accuracy (b = 59.70, SE = 9.81, t(337) = 6.09, p < 0.001, 95% CI [40.40, 78.98]) and predicted future accuracy (b = 74.66, SE = 10.38, t(337) = 7.19, p < 0.001, 95% CI [54.23, 95.08]), indicating that participants’ estimates of their own accuracy generally reflected their objective accuracy.
Yet, in the same model, reward was also significantly associated with estimated previous accuracy (b = 2.87, SE = 0.87, t(337) = 3.32, p < 0.001, 95% CI [1.17, 4.57]; Fig. 3A) and predicted future accuracy (b = 2.70, SE = 0.91, t(337) = 2.95, p = 0.003, 95% CI [0.90, 4.50]; Fig. 3B), indicating that reward exerted an influence on self-evaluations even after taking into account the effect of objective accuracy. That is, participants who received more frequent rewards for correct answers thought they had performed more accurately and would perform more accurately in the future, even when accounting for their objective performance. Specifically, participants in the high-reward condition estimated higher accuracy (M = 61.91, SE = 0.89) than in the low-reward condition (M = 57.22, SE = 0.92). Although not preregistered, a two-tailed independent samples t-test assessing the absolute difference between conditions supported our main findings (t(334.28) = 2.59, d = 0.28, p = 0.01, 95% CI [1.12, 8.25]). Participants in the high-reward condition also predicted greater accuracy (M = 65.19, SE = 0.97) than those in the low-reward condition (M = 61.11, SE = 0.98), and a (non-preregistered) two-tailed independent samples t-test also provided converging evidence for our primary regression results (t(335.32) = 2.09, d = 0.23, p = 0.04, 95% CI [0.24, 7.92]). Given that participants underestimated their own accuracy overall, those in the high-reward condition came closer to their true accuracy than those in the low-reward condition (refer to Supplementary Note 1). In the corresponding regression predicting participants’ competence ratings, the two predictors explained 6.8% of the variance (Adjusted R2 = 0.068, F(2, 337) = 13.4, f2 = 0.08, p < 0.001). In this model, accuracy significantly predicted competence ratings (b = 56.79, SE = 11.40, t(337) = 4.98, p < 0.001, 95% CI [34.37, 79.21]), but the effect of reward was non-significant (b = 1.90, SE = 1.00, t(337) = 1.89, p = 0.059, 95% CI [−0.08, 3.87]; Fig. 3C). Although participants in the high-reward condition rated higher game competence (M = 62.57, SE = 1.00) than in the low-reward condition (M = 59.78, SE = 1.07), a (non-preregistered) two-tailed independent samples t-test also did not provide evidence of a significant difference between the two conditions (t(331.44) = 1.35, d = 0.15, p = 0.20, 95% CI [−1.28, 6.87]).
In each panel, individual data points are shown for each reward condition (on the left). N = 340 participants. Difference density plots (on the right) were generated by bootstrapped resampling (5000 times) to create an empirical distribution of the difference in means between high and low-reward conditions. Circles (on left) and triangles (on right) indicate means. Error bars represent ± 1 SE from the mean. A Distribution of estimated accuracy by reward condition. B Distribution of predicted accuracy by reward condition. C Distribution of competence ratings by reward condition.
These findings replicated a pilot study that found similar effects of reward on evaluations of estimated previous accuracy and predicted future accuracy (see Supplementary Note 2). Moreover, in an exploratory (non-preregistered) analysis, we examined the relationship of self-beliefs with experienced reward rate, which may track more closely with beliefs. Although an exploratory manipulation check confirmed that participants in the high-reward condition received significantly higher average bonuses (M = 66.89 cents, SD = 8.51) than those in the low-reward condition (M = 34.50 cents, SD = 5.23, t(294.0) = 42.59, p < 0.001, d = 4.55, 95% CI [30.89, 33.88]), actual reward rates differed across individuals within the same condition due to the probabilistic nature of feedback. That is, different individuals in the same reward condition could have had relatively higher or lower reward rates compared to one another, and those different rates of reward might produce different self-beliefs. We therefore refit the linear regression models predicting task-specific self-beliefs using each participant’s actual reward rate rather than reward condition. This approach provided a more precise metric of the actual rate of reward each participant experienced. In these models, reward rate was significantly associated with estimated accuracy (b = 10.67, SE = 2.85, t(337) = 3.75, p < 0.001, 95% CI [5.07, 16.28], predicted accuracy (b = 9.45, SE = 3.02, t(337) = 3.12, p = 0.002, 95% CI [3.50, 15.40]), and competence ratings (b = 7.55, SE = 3.32, t(337) = 2.28, p = 0.024, [1.02, 14.08]; see Supplementary Note 6 and Supplementary Table 9).
Although the staircase procedure held accuracy constant, different participants might have reached 70% accuracy at different difficulty levels of the task. If difficulty levels differed across conditions, then this difference might also explain self-beliefs across conditions. We therefore tested whether the difficulty level of the staircase impacted task-specific self-belief measures. This analysis was also not originally in our preregistered plan, but we aimed to ensure results remained the same when adjusting for the objective difficulty of the game, given that participants might have been aware of task difficulty and used difficulty as a cue to their performance. To do so, we refit the regression models predicting the three task-specific self-belief measures with an additional predictor that reflects the ending difficulty level of the task for each participant. The reward effect remained significant for estimated accuracy (b = 2.89, SE = 0.87, t(336) = 3.33, p < 0.001, 95% CI [1.18, 4.59]) and predicted accuracy (b = 2.62, SE = 0.91, t(336) = 2.87, p = 0.004, 95% CI [0.82, 4.41]; see Supplementary Table 2). The average ending difficulty level of the task also did not differ significantly by condition (see Supplementary Note 4 for details). Thus, we observe no credible evidence that an awareness of the difficulty of the game could account for the observed effect of reward on self-beliefs.
While we had preregistered an exclusion rule for our analyses, we also tested whether the main findings were robust to alternative exclusion criteria. For example, our preregistered rule used trials after the main phase, but it is possible that an exclusion rule relating to performance during the main 80 rounds of the games could capture inattention better. Specifically, in exploratory analyses, we excluded participants who spent most of the game at the easiest level (i.e., no noise in dot motion). For these analyses, the inferences remained the same (see Supplementary Note 4).
Did the effects of reward extend to perceptions of the game itself, beyond self-evaluations of ability? In exploratory analyses, we examined this question by testing the effects of reward on self-reported enjoyment, difficulty, and motivation in the task using linear multiple regression. Reward did not significantly predict enjoyment (b = 2.08, robust SE = 1.48, t(337) = 1.40, p = 0.16, 95% CI [−0.83, 4.99]) or motivation (b = 0.70, SE = 1.27, t(337) = 0.55, p = 0.59, 95% CI [−1.80, 3.19]), but it did predict perceived difficulty (b = −2.50, SE = 1.23, t(337) = 0-2.03, p = 0.043, 95% CI [−4.92, −0.08]; see Supplementary Note 9 for model details). Specifically, receiving more rewards was associated with perceiving the game as less difficult. On the other hand, we observe little credible evidence that the influence of rewards on self-evaluations of ability was related to increased motivation or overall enjoyment of the game itself.
Finally, beyond overall performance judgments, we tested whether rewards during the main phase would impact later confidence when performing the task in the absence of accuracy or reward feedback. In a subsequent phase of the dot motion game, participants indicated their confidence (on a scale from 1 = Guessing to 4 = Very confident) after each trial of the game, forming a measure of local confidence. In this phase, the average accuracy was 73.60% (high reward: M = 0.74, SD = 0.14; low reward: M = 0.73, SD = 0.14), and a (non-preregistered) two-tailed independent samples t-test revealed no significant differences between reward conditions (t(333.9) = 0.72, d = 0.08, p = 0.47, 95% CI [−0.02, 0.04]). The high-reward condition rated an average local confidence across trials of 2.74 (SD = 0.71), while the low-reward condition rated an average local confidence across trials of 2.71 (SD = 0.62).
We predicted that receiving more rewards when playing the game in the first part of the experiment would lead participants to feel more confident that they would later answer correctly in the absence of any feedback. In a generalized estimating equation (GEE) model, we used accuracy on each trial and reward condition to predict trial-by-trial confidence ratings, allowing for an interaction between the two. Trial-by-trial accuracy (i.e., correct versus incorrect response) significantly predicted local confidence ratings (b = 0.06, SE = 0.005, z = 161.13, p < 0.001, 95% CI [0.05, 0.07]), indicating that participants were attuned to whether they had answered correctly. However, reward did not significantly predict local confidence overall (b = 0.0001, SE = 0.01, z = 0.00, p = 0.99, 95% CI [−0.03, 0.03]). Accuracy and reward did show a significant interaction (b = 0.01, SE = 0.005, z = 4.72, p = 0.03, 95% CI [0.001, 0.02]), suggesting that participants in the high reward (versus low reward) condition showed more differentiated confidence between correct and incorrect answers. Altogether, however, we did not observe credible evidence that rewards boosted overall later local confidence. In addition, the interaction effect suggesting sharpened metacognitive accuracy was not replicated in Study 2 and therefore should be interpreted with caution.
Study 1 showed that rewards can bias self-evaluations of ability, particularly estimations of previous accuracy and predictions of future accuracy. Although participants observed exactly how they performed on each trial, those who received more frequent rewards from the slot machine following correct answers thought they had performed better. This finding suggests that reward had an immediate effect on information processing, though we did not observe credible evidence that this effect boosted later confidence in new responses made in the absence of reward. What mechanism might support the bias in task-specific self-beliefs? Study 2 investigated the trial-by-trial influence of rewards on expectations of accuracy as people develop a sense of their ability in the game. In particular, we asked whether rewards amplify the extent to which people learn about their own ability from good performance, leading to larger updates in self-beliefs. That is, after a correct response, people can increase their estimate of their performance level, but this update may be greater when feedback is accompanied by a reward.
In Study 2, we used the same task as in Study 1 with some key additions. First, participants were now asked to indicate whether they expected to answer correctly or incorrectly before each trial of the game, allowing us to ask how rewards change perceptions of ability on a trial-by-trial basis. This procedure allowed us to test dynamic belief updating, asking how participants learned from each instance of feedback and updated their expectations accordingly. In addition, instead of seeing a slot machine, each trial was now associated with a trial-unique image representing a lottery ticket that could give a token for a correct answer or an incorrect receipt for an incorrect answer (Fig. 4).
After each round, participants saw a trial-unique image. For correct responses, this image had a green frame, and for incorrect responses, the image had a red frame. If the response was incorrect, nothing else was displayed before the next round began. If the response was correct, either a token or a circle with a slash appeared underneath the first image. In this example, object images (e.g., a ball of yarn) served as “incorrect receipts” while scene images (e.g., a wine cellar) represented lottery tickets that could give a token (similar to the slot machine in Study 1). Object and scene images shown are reprinted with permission from the authors of the original publications45,46.
We adapted reinforcement learning models to the task in order to model participants’ beliefs about their likelihood of answering correctly on each trial. We compared models with and without a reward bias on updating. Although the reward model specified in our preregistration fit best (protected exceedance probability = 1), model recovery suggested that the models were not sufficiently identifiable (that is, the competing models could not be consistently distinguished from one another when fit to simulated data). Thus, any interpretation of model selection or parameter estimates should be made with caution. We report detailed results from this computational modeling in Supplementary Note 10.
Although not included in our preregistration, we therefore conducted a mixed-effects logistic regression to predict trial-by-trial participant expectations using accuracy and reward feedback on the immediately preceding trial. This regression approach offers a more robust and interpretable framework and approximates the preregistered computational model with a learning rate of 1 (i.e., modeling how participants update their ratings based on feedback on the most recent trial). We predicted that reward feedback would increase expectations of accuracy above and beyond the influence of accuracy feedback, such that participants’ correct responses would have greater uptake into self-beliefs when accompanied by reward than not. Both accuracy (b = 0.66, SE = 0.05, z = 14.59, p < 0.001, 95% CI [0.57, 0.75]) and reward (b = 0.10, SE = 0.04, z = 2.65, p = 0.01, 95% CI [0.03, 0.18]) on the previous trial emerged as significant predictors, demonstrating that reward boosted expectations about performance beyond the effect of accuracy. Specifically, participants updated their expectations more in the positive direction when they had been rewarded for correct answers (M = 0.03, SE = 0.004) versus unrewarded (M = 0.02, SE = 0.0009; Fig. 5). As an exploratory analysis, two-tailed paired samples t-tests provided converging evidence that participants increased their trial-by-trial expectations significantly more not only after correct than incorrect responses (t(341) = 9.73, d = 0.53, p < 0.001, 95% CI [0.07, 0.10]), but also after correct-rewarded trials than correct-unrewarded trials (t(341) = 2.26, d = 0.12, p = 0.02, 95% CI [0.001, 0.03]). These results suggest that rewards shaped the extent to which people updated their expectations of future success, leading them to integrate recent accuracy feedback more into self-beliefs.
A Each dot represents one participant’s mean change in performance expectation following each type of trial. N = 342 participants. Diamonds indicate means. B Each dot represents the difference between one participant’s mean change in performance expectation following correct-rewarded trials and their mean change in performance expectation following correct-unrewarded trials. N = 342 participants. The triangle indicates the mean, and the error bars represent the 95% confidence interval around the mean.
We next examined the same task-specific self-belief measures as in Study 1. Validating the staircase procedure, objective accuracy was 69.0% (high reward: M = 0.70, SD = 0.08; low reward: M = 0.71, SD = 0.07). An exploratory two-tailed independent samples t-test did not reveal significant differences between reward conditions (t(332.21) = −1.92, d = 0.21, p = 0.06, 95% CI [−0.03, 0.0004]).
For estimations of previous accuracy, we again found that both accuracy (b = 84.64, SE = 10.03, t(339) = 8.44, p < 0.001, 95% CI [64.91, 104.37]) and reward (b = 1.71, SE = 0.77, t(337) = 2.23, p = 0.03, 95% CI [0.20, 3.22]) predicted estimates of accuracy. Thus, participants in the high-reward condition again believed they had performed more accurately compared to those in the low-reward condition. This model explained 17.3% of the variance in estimations of accuracy (Adjusted R2 = 0.173, F(2, 346) = 36.54, f 2 = 0.22, p < 0.001). However, reward was not significantly associated with either predictions of future accuracy (b = 1.41, SE = 0.85, t(337) = 1.65, p = 0.10, 95% CI [−0.27, 3.09]) or self-evaluations of competence (b = 1.12, SE = 0.91, t(337) = 1.23, p = 0.22, 95% CI [−0.67, 2.91]; see Supplementary Table 1). These results were consistent with exploratory analyses using actual reward rate instead of reward condition as a predictor (Supplementary Note 6 and Supplementary Table 10). These findings thus partially replicate the reward effects on task-specific self-beliefs observed in Study 1. Smaller effects on these measures in Study 2 may be due to trial-by-trial expectation questions drawing greater attention to accuracy, as well as more abstract and distal rewards than Study 1 (i.e., tokens that translated to raffle tickets for a lottery as opposed to a direct monetary bonus), which we consider further in the Discussion.
As in Study 1, we tested whether the results were robust to exclusion criteria. We fit exploratory (non-preregistered) models predicting each measure of task-level self-belief after excluding participants who spent most of the 80 rounds of the game at the easiest level. In these analyses, we observed stronger effects of reward on task-specific self-beliefs, across estimated accuracy (b = 2.53, SE = 0.73, t(383) = 3.47, p < 0.001, 95% CI [1.10, 3.97]), predicted accuracy (b = 2.23, SE = 0.79, t(383) = 2.84, p = 0.005, 95% CI [0.69, 3.78]), and competence ratings (b = 1.95, SE = 0.85, t(383) = 2.29, p = 0.02, 95% CI [0.28, 3.63]; see Supplementary Note 5 and Supplementary Table 7). We also refit multiple regression models with an additional predictor that captures the ending difficulty level of the task for each participant, and the reward effect on estimated accuracy remained significant (b = 1.68, SE = 0.77, t(336) = 2.19, p = 0.03, 95% CI [0.17, 3.20]; see Supplementary Note 4 and Supplementary Table 3).
As in Study 1, we also tested whether rewards during the main phase of the game would affect later confidence in a separate phase without accuracy or reward feedback. The average accuracy in these trials was 73.40% (high reward: M = 0.72, SD = 0.13; low reward: M = 0.74, SD = 0.14), and a (non-preregistered) two-tailed independent samples t-test revealed no significant differences between reward conditions (t(339.34) = −1.34, d = 0.14, p = 0.18, 95% CI [−0.05, 0.01]). The high-reward condition rated an average local confidence across trials of 2.86 (SD = 0.65), while the low-reward condition rated an average local confidence across trials of 2.91 (SD = 0.57).
Although trial-by-trial accuracy significantly predicted local confidence (b = 0.05, SE = 0.004, z = 175.20, p < 0.001, 95% CI [0.05, 0.06]), reward did not (b = −0.005, SE = 0.01, z = 0.18, p = 0.67, 95% CI [−0.03, 0.02]). The interaction between the two was also non-significant (b = −0.004, SE = 0.004, z = 1.07, p = 0.30, 95% CI [−0.01, 0.004]). The interaction was somewhat sensitive to model assumptions, such that an alternative model specification suggested a negative interaction between reward and accuracy (b = −0.02, SE = 0.002, z = −11.80, p < 0.001, 95% CI [−0.03, −0.02]), unlike Study 1 (see Supplementary Note 8 for details). Overall, we observe no credible evidence that rewards in the main phase of the game boosted local confidence in the absence of feedback, and we did not observe consistent interaction effects across studies.
We again examined whether rewards influenced perceptions of the game beyond self-evaluations of ability. In multiple regression models, reward was not significantly associated with enjoyment (b = −0.21, SE = 1.36, t(339) = −0.15, p = 0.88, 95% CI [−2.90, 2.47]), perceived difficulty (b = −1.60, SE = 1.15, t(339) = −1.39, p = 0.17, 95% CI [−3.86, 0.66]), or self-reported motivation (b = −0.64, SE = 1.32, t(339) = −0.49, p = 0.63, 95% CI [−3.22, 1.93]; see Supplementary Note 9 and Supplementary Table 14).
Finally, in order to assess links to broader self-views, we added a measure of general self-esteem in Study 2 and examined the link between this measure and task-specific self-beliefs. In a multiple linear regression model with reward and self-esteem as predictors, self-esteem was significantly associated with competence ratings (b = 3.37, SE = 1.43, t(314) = 2.36, p = 0.02, 95% CI [0.55, 6.19]), which reflect generalized judgments about ability in the task, though not with estimations of prior accuracy (b = 1.85, SE = 1.21, t(314) = 1.52, p = 0.13, 95% CI [−0.54, 4.24]) or predictions of future accuracy (b = 1.66, SE = 1.36, t(314) = 1.22, p = 0.22, 95% CI [−1.02, 4.33]). These results suggest a link between task-specific competence judgments and general views about the self (see Supplementary Note 9 and Supplementary Tables 15 and 16 for more details).
Across two behavioral experiments, we examined the impact of rewards on self-evaluations of ability in a task by decoupling rewards from objective performance feedback. Results showed that rewards biased people’s understanding of their prior performance, even when they had access to objective performance feedback. We then probed the learning dynamics behind this bias, finding that rewards increase the extent to which people update their beliefs following good performance in a task.
These findings expand our understanding of the links between reward learning and self-beliefs. Prior research has already established several links between rewards and self-beliefs. First, rewards influence motivation, which can shape future performance, and, in turn, impact self-beliefs8,50,51,52. Second, in many cases, rewards are reflective of ability to some extent, and in these situations, people may use rewards as diagnostic indicators of performance to inform self-evaluations. The current findings extend prior work by demonstrating that rewards may still exert a direct influence on self-evaluations beyond motivation or any informative value, potentially by virtue of their affective qualities. This account would be in line with prior work showing affect-related biases in forming and updating self-beliefs14,53 and reward-driven biases in belief updating more generally54,55. Our findings also complement recent work showing that rewards bias people’s impressions of others’ competence even when objective performance is known20, and work demonstrating that monetary incentives for accurate confidence judgments biased these very judgments18. The present work demonstrates that reward biases people’s overall impressions of their own performance and abilities in a task domain, even when accuracy feedback is provided. In contrast, we did not detect consistent evidence that rewards for correct responses influenced metacognitive confidence in later responses in the absence of reward.
The present studies also add to a body of work on inaccuracies in people’s self-evaluations of ability. Perceptions of skill often show only small correlations with objective performance56. The underestimation of performance in the current work aligns with previous work showing that people update their estimates of their performance more after negative feedback when they perceive an opportunity to improve, as when learning a new skill28. Given that self-evaluations then shape goal persistence and future outcomes1, understanding the nature of these flawed self-assessments can help inform interventions to mitigate their detrimental effects. Here, we suggest an additional way in which self-beliefs can fail to reflect reality: They can be constructed through the experience of reward, which may reflect influences beyond objective performance, such as the availability of resources within an institution.
Although our current studies afforded several benefits pertaining to experimental control, they are also limited in generalizability to more complex, real-world learning scenarios. By using perceptual games, we presented participants with a task in which they were unlikely to have an existing self-concept. This approach also allowed us to use a staircase procedure to control average performance, ensuring that differences in reward did not promote differences in rates of correct responses between conditions on average. On the other hand, learning a new skill in the real world is likely to last much longer than the length of our current experiments and typically occurs over many repeated experiences. Moreover, the rewards used in the present studies were small; in Study 1, tokens translated to monetary bonuses of under a dollar, and Study 2 used an indirect reward structure, in which tokens translated to raffle tickets in a lottery for a 10-dollar gift card. These reward manipulations may blunt the positive feelings that accompany rewards in real life. The change to a less direct reward structure in Study 2 may have contributed to the smaller effect sizes observed in Study 2 compared to Study 1 on the task-specific self-belief measures. In addition, drawing more attention to trial-by-trial accuracy in Study 2 may have dampened the influence of reward on overall beliefs as compared to Study 1 and deviated more from the way in which people spontaneously track their own ability. Nonetheless, this approach allowed us to assess moment-to-moment changes in self-beliefs during learning as opposed to retroactive judgment, informing how rewards change information uptake.
In the present study, rewards were extrinsic. This approach was used to dissociate performance and reward, such that correct performance could be rewarded or unrewarded. However, similar psychological processes may apply to intrinsic reward. Good performance can be intrinsically rewarding and elicit feelings of pride11, and there is overlap in neural representation of reward value, accuracy, and confidence22,25. Even in these cases, reward could have an affective impact that boosts self-evaluations of ability above and beyond its informative properties. Indeed, beliefs about the self tend to hold intrinsic value and induce emotions that can bias learning10,14. Accordingly, the current results do not necessarily suggest that extrinsic reward is needed to generate self-confidence; in fact, extrinsic reward can also have negative effects on long-term motivation57. Rather, they highlight a direct role of reward in shaping self-beliefs. Future work can test how intrinsic affective responses relate to these shifts in self-beliefs.
In the current study, rewards were delivered probabilistically (via a slot machine or lottery ticket), but only after correct answers. It therefore remains an open question how the observed reward bias may change in situations when reward feedback is either entirely unrelated to accuracy or even more tightly linked to performance. When it is clear exactly why a reward was or was not garnered for a particular response, it is possible that people can more easily determine whether they earned it, which might shape the extent to which rewards influence self-beliefs. Future work could examine how uncertainty around rewards affects the impact of feedback on self-perceptions. In addition, the current study used extrinsic monetary rewards, but rewards also often come in the form of praise or verbal reinforcement. Future work can test the role of social rewards, which rely on similar brain processes to monetary rewards58 but may have different effects on intrinsic motivation59.
Nonetheless, given that people often receive extrinsic rewards for performance in the real world, future work can test how the current findings may apply to scholastic achievement. Students’ self-beliefs directly influence performance and future achievement, informing their decisions about which challenges to take on and what paths to pursue to maximize chances of success60.
In sum, we identify an influence of rewards on self-belief formation, above and beyond objective feedback about performance. This work illuminates the computations through which people learn about their abilities over time and highlights influences from reward learning that can complement more conceptual forms of learning61. By dissociating affective influences from objective information, these findings can inform how individuals form expectations of success that guide significant decisions about when to persevere and when to give up.
The raw, trial-by-trial subject-level data that support the current findings are available in the study’s online repository in CSV format: https://doi.org/10.17605/OSF.IO/Z4J9862.
The code for running the experiment (Psychopy file and dependencies) and for all analyses of the data (R files) that appear in both the main text and supplemental information can be found in the study’s online repository: https://doi.org/10.17605/OSF.IO/Z4J9862.
Fouad, N. A., Smith, P. L. & Zao, K. E. Across academic domains: extensions of the social-cognitive career model. J. Counsel. Psychol. 49, 164–171 (2002).
Article Google Scholar
Oyserman, D. Identity-based motivation. Emerg. Trends Soc. Behav. Sci. 38, 1–11 (2015).
Markus, H. & Wurf, E. The dynamic self-concept: a social psychological perspective. Annu. Rev. Psychol. 38, 299–337 (1987).
Article Google Scholar
Markus, H. & Kunda, Z. Stability and malleability of the self-concept. J. Personal. Soc. Psychol. 51, 858–866 (1986).
Article Google Scholar
Gore, J. S. & Cross, S. E. Who am i becoming? A theoretical framework for understanding self-concept change. Self Identity 13, 740–764 (2014).
Article Google Scholar
Condly, S. J., Clark, R. E. & Stolovitch, H. D. The effects of incentives on workplace performance: a meta-analytic review of research studies. Perform. Improv. Q. 16, 46–63 (2003).
Article Google Scholar
Jenkins Jr, G. D., Mitra, A., Gupta, N. & Shaw, J. D. Are financial incentives related to performance? A meta-analytic review of empirical research. J. Appl. Psychol. 83, 777 (1998).
Article Google Scholar
Bem, D. J. Advances in Experimental Social Psychology, Vol. 6. (ed. Berkowitz, L.) 1–62 (Academic Press, 1972).
Melnikoff, D. E. & Strohminger, N. Bayesianism and wishful thinking are compatible. Nat. Hum. Behav. 8, 692–701 (2024).
Article PubMed Google Scholar
Bromberg-Martin, E. S. & Sharot, T. The value of beliefs. Neuron 106, 561–565 (2020).
Article PubMed Google Scholar
Müller-Pinzler, L. et al. Neurocomputational mechanisms of affected beliefs. Commun. Biol. 5, 1241 (2022).
Article PubMed PubMed Central Google Scholar
Storbeck, J. & Clore, G. L. Affective arousal as information: How affective arousal influences judgments, learning, and memory. Soc. Personal. Psychol. Compass 2, 1824–1843 (2008).
Article PubMed PubMed Central Google Scholar
Koban, L. et al. Social anxiety is characterized by biased learning about performance and the self. Emotion 17, 1144–1155 (2017).
Article PubMed PubMed Central Google Scholar
Sharot, T. & Garrett, N. Forming beliefs: why valence matters. Trends Cogn. Sci. 20, 25–33 (2016).
Article PubMed Google Scholar
Schwarz, N. Handbook of Theories of Social Psychology Volume 1 (eds Van Lange, P. A. M., Kruglanski, A. & Higgins, E. T.) 289–308 (SAGE Publications Ltd, 2012).
Schwarz, N. & Clore, G. L. Mood, misattribution, and judgments of well-being: Informative and directive functions of affective states. J. Personal. Soc. Psychol. 45, 513–523 (1983).
Article Google Scholar
Song, H. & Schwarz, N. If it’s hard to read, it’s hard to do: processing fluency affects effort prediction and motivation. Psychol. Sci. 19, 986–988 (2008).
Article PubMed Google Scholar
Lebreton, M. et al. Two sides of the same coin: monetary incentives concurrently improve and bias confidence judgments. Sci. Adv. 4, eaaq0668 (2018).
Article PubMed PubMed Central Google Scholar
Rouault, M., Dayan, P. & Fleming, S. M. Forming global estimates of self-performance from local confidence. Nat. Commun. 10, 1141 (2019).
Article PubMed PubMed Central Google Scholar
Hackel, L. M., Kogon, D., Amodio, D. M. & Wood, W. Group value learned through interactions with members: a reinforcement learning account. J. Exp. Soc. Psychol. 99, 104267 (2022).
Article Google Scholar
Daniel, R. & Pollmann, S. Striatal activations signal prediction errors on confidence in the absence of external feedback. NeuroImage 59, 3457–3467 (2012).
Article PubMed Google Scholar
Foerde, K. & Shohamy, D. Feedback timing modulates brain systems for learning in humans. J. Neurosci. 31, 13157–13167 (2011).
Article PubMed PubMed Central Google Scholar
Rouault, M. & Fleming, S. M. Formation of global self-beliefs in the human brain. Proc. Natl Acad. Sci. USA 117, 27268–27276 (2020).
Article PubMed PubMed Central Google Scholar
Fleming, S. M., Huijgen, J. & Dolan, R. J. Prefrontal contributions to metacognition in perceptual decision making. J. Neurosci. 32, 6117–6125 (2012).
Article PubMed PubMed Central Google Scholar
Hoven, M. et al. Motivational signals disrupt metacognitive signals in the human ventromedial prefrontal cortex. Commun. Biol. 5, 244 (2022).
Article PubMed PubMed Central Google Scholar
Wittmann, M. K. et al. Self-other mergence in the frontal cortex during cooperation and competition. Neuron 91, 482–493 (2016).
Article PubMed PubMed Central Google Scholar
Wittmann, M. K. et al. Causal manipulation of self-other mergence in the dorsomedial prefrontal cortex. Neuron 109, 2353–2361.e11 (2021).
Article PubMed PubMed Central Google Scholar
Müller-Pinzler, L. et al. Negativity-bias in forming beliefs about own abilities. Sci. Rep. 9, 14416 (2019).
Article PubMed PubMed Central Google Scholar
Fleming, S. M. Metacognition and confidence: a review and synthesis. Annu. Rev. Psychol. 75, 241–268 (2024).
Article PubMed Google Scholar
Faul, F., Erdfelder, E., Buchner, A. & Lang, A. Statistical power analyses using g* power 3.1: tests for correlation and regression analyses. Behav. Res. Methods 41, 1149–1160 (2009).
Article PubMed Google Scholar
Peirce, J. W. Psychopy-psychophysics software in Python. J. Neurosci. Methods 162, 8–13 (2007).
Article PubMed PubMed Central Google Scholar
Leek, M. R. Adaptive procedures in psychophysical research. Percept. Psychophys. 63, 1279–1292 (2001).
Article PubMed Google Scholar
Kuroki, D. & Pronk, T. jsquestplus: a Javascript implementation of the quest+ method for estimating psychometric function parameters in online experiments. Behav. Res. Methods 55, 3179–3186 (2023).
Article PubMed Google Scholar
Watson, A. B. & Pelli, D. G. Quest: a Bayesian adaptive psychometric method. Percept. Psychophys. 33, 113–120 (1983).
Article PubMed Google Scholar
Rahnev, D. & Fleming, S. M. How experimental procedures influence estimates of metacognitive ability. Neurosci. Conscious. 2019, niz009 (2019).
Article PubMed PubMed Central Google Scholar
Fleming, S. M. & Lau, H. C. How to measure metacognition. Front. Hum. Neurosci. 8, 443 (2014).
Article PubMed PubMed Central Google Scholar
R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/ (R Core Team, 2024).
Wickham, H. ggplot2: Elegant Graphics for Data Analysis. https://ggplot2.tidyverse.org (Springer-Verlag New York, 2016).
Kay, M. ggdist: Visualizations of Distributions and Uncertainty). https://mjskay.github.io/ggdist/. R package version 3.3.2 (2024).
Wilke, C. O. cowplot: Streamlined Plot Theme and Plot Annotations for ‘ggplot2’. https://CRAN.R-project.org/package=cowplot. R package version 1.1.3 (2024).
Zeileis, A. Object-oriented computation of sandwich estimators. J. Stat. Softw. 16, 1–16 (2006).
Article Google Scholar
Zeileis, A. & Hothorn, T. Diagnostic checking in regression relationships. R. N. 2, 7–10 (2002).
Google Scholar
Yan, J. geepack: Yet another package for generalized estimating equations. R.-N. 2/3, 12–14 (2002).
Google Scholar
Christensen, R. H. B. Ordinal—Regression Models for Ordinal Data. https://CRAN.R-project.org/package=ordinal. R package version 2023.12-4.1 (2023).
Konkle, T., Brady, T. F., Alvarez, G. A. & Oliva, A. Conceptual distinctiveness supports detailed visual long-term memory for real-world objects. J. Exp. Psychol.: Gen. 139, 558 (2010).
Article PubMed Google Scholar
Konkle, T., Brady, T. F., Alvarez, G. A. & Oliva, A. Scene memory is more detailed than you think: the role of categories in visual long-term memory. Psychol. Sci. 21, 1551–1556 (2010).
Article PubMed Google Scholar
Rosenberg, M. Rosenberg self-esteem scale (rse). Accept. Commit. Ther. Meas. Package 61, 18 (1965).
Google Scholar
Kroenke, K., Spitzer, R. L. & Williams, J. B. The PHQ-9: validity of a brief depression severity measure. J. Gen. Intern. Med. 16, 606–613 (2001).
Article PubMed PubMed Central Google Scholar
Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67, 1–48 (2015).
Article Google Scholar
Skinner, B. F. The generic nature of the concepts of stimulus and response. J. Gen. Psychol. 12, 40–65 (1935).
Article Google Scholar
Hendijani, R., Bischak, D. P., Arvai, J. & Dugar, S. Intrinsic motivation, external reward, and their effect on overall motivation and performance. Hum. Perform. 29, 251–274 (2016).
Article Google Scholar
Eisenberger, R., Pierce, W. D. & Cameron, J. Effects of reward on intrinsic motivation-negative, neutral, and positive: comment on Deci, Koestner, and Ryan (1999). Psychol. Bull. 125, 677–691 (1999).
Article PubMed Google Scholar
Koellinger, P. & Treffers, T. Joy leads to overconfidence, and a simple countermeasure. PLoS ONE 10, e0143263 (2015).
Article PubMed PubMed Central Google Scholar
Babür, B. G., Leong, Y. C., Pan, C. X. & Hackel, L. M. Neural responses to social rejection reflect dissociable learning about relational value and reward. Proc. Natl Acad. Sci. USA 121, e2400022121 (2024).
Article PubMed PubMed Central Google Scholar
Fischer, A. G., Bourgeois-Gironde, S. & Ullsperger, M. Short-term reward experience biases inference despite dissociable neural correlates. Nat. Commun. 8, 1690 (2017).
Article PubMed PubMed Central Google Scholar
Dunning, D., Heath, C. & Suls, J. M. Flawed self-assessment: implications for health, education, and the workplace. Psychol. Sci. Public Interest 5, 69–106 (2004).
Article PubMed Google Scholar
Cameron, J. Negative effects of reward on intrinsic motivation-a limited phenomenon: comment on Deci, Koestner, and Ryan (2001). Rev. Educ. Res. 71, 29–42 (2001).
Article Google Scholar
Bhanji, J. P. & Delgado, M. R. The social brain and reward: social information processing in the human striatum. WIREs Cogn. Sci. 5, 61–73 (2014).
Article Google Scholar
Cameron, J. & Pierce, W. D. Reinforcement, reward, and intrinsic motivation: a meta-analysis. Rev. Educ. Res. 64, 363–423 (1994).
Article Google Scholar
Felson, R. B. The effect of self-appraisals of ability on academic performance. J. Personal. Soc. Psychol. 47, 944–952 (1984).
Article Google Scholar
Elder, J., Davis, T. & Hughes, B. L. Learning about the self: motives for coherence and positivity constrain learning from self-relevant social feedback. Psychol. Sci. 33, 629–647 (2022).
Article PubMed Google Scholar
Luo, J., Mende-Siedlecki, P. & Hackel, L. M. Rewards Shape Self-evaluations of Ability. Open Sci. Framework. osf.io/z4j98 (2025).
Download references
We would like to thank members of the Social Learning and Choice Lab for helpful comments and feedback.
Department of Psychology, University of Southern California, Los Angeles, CA, USA
Jean Luo & Leor M. Hackel
Department of Psychological and Brain Sciences, University of Delaware, Newark, DE, USA
Peter Mende-Siedlecki
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
J.L., L.H., and P.M.S. conceptualized the study. J.L. programmed the task, collected the data, conducted statistical analyses, and drafted the manuscript. L.H. and P.M.S. supported analyses, provided supervision, offered critical feedback on drafts, and approved the final manuscript.
Correspondence to Jean Luo.
The authors declare no competing interests.
Communications Psychology thanks Niv Reggev and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary handling editors: Troby Ka-Yan Lui. [A peer review file is available].
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Luo, J., Mende-Siedlecki, P. & Hackel, L.M. Rewards bias self-evaluations of ability. Commun Psychol 3, 143 (2025). https://doi.org/10.1038/s44271-025-00286-7
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s44271-025-00286-7
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Communications Psychology (2025)
Advertisement
Communications Psychology (Commun Psychol)
ISSN 2731-9121 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.