Individuals who are socially anxious tend to see the world through rejection-colored glasses, perceiving ambiguous social situations as more negative than they are and expecting to perform poorly in social situations [1,2]. Some of these negative cognitive biases might be partly maintained by negative learning biases, wherein individuals who are socially anxious have difficulty using positive information to update their previously held negative beliefs [3]. For example, if a person who is socially anxious updated their beliefs about an acquaintance more from the few times that the acquaintance seemed annoyed with them than from the many times the acquaintance was friendly, they might come to have an overly negative expectancy of interactions with that acquaintance. Similarly, negative expectancies about social performance might be partially explained by overweighting instances of negative versus positive feedback—learning more from criticism than praise. Changing these learning processes to be less negatively biased may have positive effects for individuals who are socially anxious, such as changing rigidly negative beliefs about social interactions and performance to be more balanced and reducing motivation to avoid social situations. In this study, we investigated whether a computerized intervention that trains more positive, flexible interpretation styles for individuals who are socially anxious also modifies the learning process by which they update their social expectancies from new experiences (ie, social reinforcement learning [RL]). We used computational modeling to understand changes in particular parameters of interest, specifically learning rates, which are the weights given to positive and negative feedback when updating social expectancies.

RL is the process by which people or other agents learn to predict outcomes and optimize behavior in an environment where taking actions leads to rewards (positive outcomes) and punishments (negative outcomes) [4]. RL has been used to understand various mental illnesses, including addiction [5], depression [6], and psychosis [7], and has been used to develop algorithms for diverse applications, such as self-driving cars [8], health care decisions [9], and gaming [10]. Social RL is the same learning process but specific to the social environment: updating beliefs about different social actions based on social rewards (such as acceptance and smiles from others) and social punishments (such as rejection and scowls). Previous research has begun to explore social RL in social anxiety, finding that, unlike healthy individuals, who update their beliefs about their social performance more from positive versus negative feedback provided by others, individuals who are socially anxious update more from negative feedback [11]. When confronted with others who evaluate them negatively or neutrally, people who are socially anxious readily learn these evaluations accurately, whereas healthy controls are biased toward believing that others think more positively of them even when they are receiving more negative or neutral evaluations [12]. Some research has found that individuals who are socially anxious tend to learn more from angry versus happy faces [13], whereas other research has not found evidence that people who are socially anxious have biased updating of angry faces [14]. Importantly, several studies have found that individuals who are socially anxious have difficulty adapting their learning and behavior to changing probabilities of social punishment [15-17]. Specifically, there is accumulating evidence that individuals who are socially anxious have difficulty updating negative social expectancies to become more positive [14,18]. Together, these studies suggest that individuals who are socially anxious may have negative biases in social RL that can make it difficult to learn from social interactions, even when they seemingly go well.

Although there is considerable literature on the neural underpinnings of RL and how pharmacological interventions may change RL, less is known about changing RL through cognitive and behavioral interventions. Several pharmacological interventions have been used to normalize aberrant processing of rewards and punishments in various psychological disorders [19-22], including specifically changing punishment learning rates (ie, shifting how heavily a person weighs a negative outcome when updating expectancies) [23,24]. However, very little research has assessed whether cognitive behavioral interventions change RL processes, despite several researchers advocating for computationally modeling learning and cognitive change in cognitive behavioral therapy [25,26]. Recent studies have shown that cognitive interventions can change RL processes in healthy individuals [27] and individuals with schizophrenia [28]. To our knowledge, this study is the first to behaviorally assess changes in RL through cognitive bias modification in a population with elevated anxiety symptoms.

This study assesses whether several social RL biases, which are biases in learning from positive and negative social feedback to update beliefs about others and one’s own social performance, can be changed through a scalable, computerized intervention called cognitive bias modification for interpretations (CBM-I). CBM-I programs aim to train people to make less threatening and more flexible interpretations of ambiguous stimuli. They often do this by presenting scenarios that are emotionally ambiguous until the final word. The final word more frequently resolves the ambiguity in positive or neutral ways than negative ways, which aims to train participants to interpret ambiguous situations more benignly. CBM-I is a targeted, mechanism-driven intervention that is easily scalable. Because individuals who are socially anxious can use this intervention from the comfort of their own homes or wherever it is convenient for them, CBM-I holds potential either as an adjunct to other treatments or as a stand-alone intervention. CBM-I can effectively reduce cognitive biases and, to a lesser extent, the symptoms of social anxiety disorder [29-32]. However, there has been mixed evidence regarding the efficacy of CBM-I, including small effect sizes [33,34].

CBM-I is a useful intervention to study changes in social RL because the cognitive mechanism is conceptually related. Training people to interpret ambiguous situations more benignly may facilitate changes in social RL by priming awareness of potential social rewards and reducing expectations of social rejection. For example, in the CBM-I intervention used in this study, participants read emotionally ambiguous scenarios in which the ambiguity is only resolved in the final word, which is missing a letter that participants complete (eg, “You are required to give a presentation to a group of work colleagues that you know well. They all are quiet during your presentation. As you think about the presentation later at home, you think that your colleagues found your presentation...stimul_ting”). As a participant who is socially anxious reads this scenario, they might start to think the scenario will end poorly because they expect punishment in socially threatening situations (eg, that their colleagues hated their presentation), but the actual resolution of the ambiguity in a rewarding way (ie, that their colleagues found their presentation stimulating) might train the participant to update their future expectancies to be less threatening. Notably, previous work has found that RL parameters predict the response to cognitive bias modification to shift attention bias (ie, selective focus on threat cues) [35]; given their similar focus on shifting threat biases, CBM-I may also be related to RL.

Three previous studies [36,37] (Beltzer, ML, unpublished data, June 2023) that are relevant to this study have been performed using data from the same data set, and these studies informed our hypotheses. Two of these studies [36] (Beltzer, ML, unpublished data, June 2023) assessed social anxiety–related differences in social RL by comparing participants with high social anxiety symptoms (who had not completed CBM-I) with participants with low social anxiety symptoms on 2 social RL tasks. First, the social probabilistic selection task assessed learning about other people by asking participants to choose between neutral faces with different probabilities of becoming happy or angry when chosen. Second, the speech expectancies task assessed learning about one’s own social performance by asking participants to rate how they expected to perform on a stressful speech both before the speech and again after the speech and seeing feedback from the judges on their performance (the tasks are fully described in the Methods section and in their original papers [36] [Beltzer, ML, unpublished data, June 2023]). For each of these tasks, we used RL to computationally model how each participant weighted rewarding and punishing social outcomes when updating their beliefs, parameterized as the reward and punishment learning rates. These studies did not find social anxiety–related differences in learning rates as a function of social anxiety group or whether social feedback was rewarding or punishing but did find that, contrary to hypotheses, participants with high versus low social anxiety symptoms were less accurate at avoiding punishing faces once feedback was no longer given [36]. Exploratory analyses found that participants (regardless of social anxiety group) updated their expectancies of their social performance more from positive than negative feedback and that participants with high versus low social anxiety symptoms used feedback to update their expectancies more for elements of poor social performance than good social performance (Beltzer, ML, unpublished data, June 2023). These results suggest that negative social RL biases were not evident in this sample in the ways we had expected; instead, participants with high social anxiety showed impaired punishment learning accuracy and greater updating from feedback about feared, negative aspects of social performance.

Another study from this data set [37] assessed the effects of CBM-I on outcomes related to cognitive styles and social anxiety symptoms. Participants who completed both CBM-I and ecological momentary assessment (EMA) about their emotion regulation experienced greater reductions in trait negative interpretation bias than participants who completed only EMA. However, CBM-I was not uniquely related to changes in other trait cognitive styles (ie, cognitive flexibility, reappraisal ability, social interaction anxiety, and fear of negative evaluation). CBM-I also increased the daily self-reported ability to use cognitive reappraisal but it did not change other daily cognitive style outcomes more than only EMA. These results suggest that, in this data set, the CBM-I intervention effectively changed negative interpretation biases but did not have effects on many other processes related to social anxiety that it did not target as directly. Given social RL’s close conceptual ties to interpretation biases, it is possible that although null effects were found on many of these peripheral outcomes, CBM-I might still exert effects on social RL.

This study examined whether social RL biases changed as a function of completing a week of CBM-I in the middle of a 5-week EMA study tracking emotion regulation strategy use and affect (as compared with the comparison condition that also completed 5 weeks of EMA but no CBM-I). We assessed whether CBM-I was associated with differences in the weights given to socially rewarding versus punishing information when updating beliefs in 2 domains: learning about other people and learning about one’s own social performance. Given the slight methodological differences, these are referred to as “learning rates” for learning about other people and “update weights” for updating social performance expectancies based on feedback. In the domain of learning about other people, we also assessed whether CBM-I was associated with differences in using what one has learned about others to guide one’s own behavior to accurately choose rewarding people and to avoid punishing people. To do this, we compared parameters extracted from a social probabilistic learning task administered at baseline and follow-up laboratory sessions, and a speech expectancy updating task administered at follow-up, as a function of the intervention group (CBM-I vs only EMA).

To our knowledge, no study to date has assessed changes in RL as a function of CBM-I. We had competing hypotheses for most outcomes, as prior analyses of social RL in this data set found results that diverged from hypotheses based on extant literature. Plans for analyses and competing hypotheses were preregistered on the web [38].

On the basis of conceptually related findings of decreased negative interpretation bias after completing a course of CBM-I (observed both in the literature and in this data set, as described earlier) [37], we hypothesized that participants who are socially anxious would exhibit a decrease in parameters related to social punishment learning (punishment learning rate and accuracy in avoiding punishment) and/or an increase in parameters related to social reward learning (reward learning rate and accuracy in choosing reward). Modeling these parameters separately allowed us to pinpoint whether CBM-I led to changes in learning from social reward and/or punishment and whether those changes were reflected in the weight given to new social information and/or the accuracy of decisions made based on information learned about reward and/or punishment. Similarly, we hypothesized that socially anxious participants who completed CBM-I would update their beliefs about their own social performance more based on positive feedback and/or less based on negative feedback, as compared with the EMA-only group.

On the basis of our prior analyses on the social probabilistic learning task, finding no differences in learning rates as a function of the social anxiety group, we hypothesized that we would not find a difference between the intervention groups in learning rates at either session. On the basis of our surprising finding of impaired accuracy in avoiding punishment in the high social anxiety group [36], we hypothesized that CBM-I might mitigate this bias, evidenced as a greater increase in accuracy in avoiding punishment from baseline to follow-up in the CBM-I versus EMA-only group. On the basis of our prior exploratory finding of greater updating on items assessing poor social performance in the high versus low social anxiety group (Beltzer, ML, unpublished data, June 2023), we hypothesized that a model with 2 update weights (estimated separately for items assessing good vs poor social performance) would reveal a CBM-I versus EMA-only intervention–related difference. Specifically, we hypothesized that CBM-I would mitigate this bias toward greater updating on items assessing poor social performance, such that the CBM-I versus EMA-only group would update their expectancies less from feedback on items measuring poor versus good performance.

All participants completed a baseline session in the laboratory that included informed consent, questionnaires, and social probabilistic selection and speech expectancy tasks. Participants then completed 5 weeks of EMA about their emotion regulation (for more information on EMA procedures, see the studies by Daniel et al [37], Beltzer et al [40], and Daniel et al [41]), followed by another laboratory session with the same tasks (because the mobile app used for only EMA ran on certain versions of iOS and Android OS, participants who did not have a compatible smartphone were excluded. More information about the EMA component of the study, as well as other measures administered for the larger parent study but not analyzed in this study, can be found on the web [42]). Approximately half (59/114, 51.8%) of the sample was randomized to complete a week-long CBM-I intervention during week 3, and this CBM-I condition’s EMA prompts were slightly reduced during this week to ensure that the time spent on tasks was similar across conditions.

This study was approved by the University of Virginia Institutional Review Board (UVA IRB 2270). and was conducted in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and the Helsinki Declaration of 1975, as revised in 2000.

All the participants provided informed consent. Participants were compensated US $25 for each laboratory session they completed and up to US $70 for completing EMA surveys and CBM-I sessions, with more compensation for higher completion, as detailed in the informed consent form, for a total compensation of up to US $120 for the completion of all study components. Participants who were undergraduate students at University of Virginia could alternatively claim up to 4 hours of research credits for participation. Data were deidentified for analysis. Patients were fully debriefed at the end of the study. As described earlier, the feedback and expectancy updating portion of the speech expectancies task was administered only at each participant’s final session to be able to debrief about false feedback on the same day they received feedback.

Participants’ behavior on the social probabilistic selection task was analyzed to better understand how CBM-I affected social RL about other people. Specifically, we assessed reward and punishment learning rates, which described the weights given to happy and angry faces, respectively, in updating expectancies of other people during the training phase and accuracy in choosing rewarding and avoiding punishing faces in the testing phase.

This study assessed the effects of CBM-I on social RL in individuals who are socially anxious, focusing on 2 important domains: learning about other people and learning about one’s own social performance. For the most part, we did not find evidence supporting the 3-way interactions that we had hypothesized (ie, interactions among session, intervention condition, and prediction error valence), but we did find evidence of several related 2-way interactions. After versus before CBM-I, participants updated their expectancies of other people on the social probabilistic selection task more slowly from new social information and used what they learned to make more accurate decisions. These changes were not observed in the EMA-only condition, suggesting that they were likely related to the CBM-I intervention. Furthermore, participants who completed CBM-I used social feedback to update their expectancies of their own performance on the speech expectancies task in less-biased ways than EMA-only participants, in that they learned similarly from positive and negative feedback and learned similarly from feedback on items related to poor and good social performance. We also observed several effects as a function of session (ie, change over time), regardless of the intervention condition, which might be a result of either practice with the social probabilistic selection task or from the 5 weeks of EMA tracking affect and emotion regulation that all participants completed. Specifically, at the end of the study, participants updated their expectancies about others more from reward and less from punishment, and they became more accurate at avoiding punishment, as compared to at the start of the study. Although the results were mixed, these findings point to several intriguing changes in social RL associated with CBM-I.

We assessed how individuals who are socially anxious update their beliefs based on new social information by analyzing learning rates on the social probabilistic selection task (to measure updating expectancies about other people) and update weights on the speech expectancies task (to measure updating expectancies about one’s own social performance). There were 3 models evaluating updating: 1 for learning about other people and 2 for learning about one’s own performance (because update weights were estimated in 2 ways).

In addition to assessing the effects of CBM-I on how individuals who are socially anxious use different types of social information to update their expectancies, we also analyzed how individuals who are socially anxious use what they have learned about other people to make decisions. This process was only measured in the domain of learning about other people and not learning about one’s own social performance. Although we did not find evidence of the hypothesized 3-way interaction, we found evidence of all the composite 2-way interactions. First, for the CBM-I condition, but not the only EMA condition, participants’ decision accuracy improved over the course of the study, such that they were able to more frequently select the more rewarding face during testing phase trials that included the most rewarding and most punishing faces. This suggests that CBM-I may improve social decision-making in individuals who are socially anxious, which may manifest in choosing to interact with people who are more likely to respond positively to them or in choosing actions that are more likely to elicit a positive response. Second, participants became better at avoiding social punishment but not at choosing social reward over the course of the study. This effect was not specific to CBM-I and suggests that either EMA about emotion regulation or practice with the task might improve accuracy in learning to avoid social punishment. This finding is particularly noteworthy because previous research, both in this data set [36] and elsewhere [17], has found that individuals who are anxious may show impairments in learning to avoid people who respond negatively to or take advantage of them. Frequent monitoring of emotions (as occurred in the EMA portion of this study) might help participants with social anxiety become more aware of when certain people make them feel bad, which might help them learn to avoid these types of punishing people.

To the best of our knowledge, this study is the first to examine the effects of CBM-I and EMA about emotion regulation on RL. Both CBM-I and EMA are digital interventions that can be scaled up at low cost to increase access to treatment, which may help bridge the treatment gap for individuals who are socially anxious. We found that CBM-I was associated with several changes in social RL that might be helpful for individuals with social anxiety. Participants who completed CBM-I were slower to change their beliefs about others based on each instance of facial feedback. In real life, this might mean that CBM-I could help people with social anxiety resist the urge to jump to a conclusion about a person based on a single bad or good interaction. CBM-I also improved the accuracy of social decision-making about other people (as evidenced by their more frequent selection of more rewarding faces in the social probabilistic selection task at follow-up vs baseline). This higher accuracy might help individuals with social anxiety navigate social situations more successfully, resulting in more positive interactions with others and better relationships. Participants who completed the CBM-I intervention also used feedback to update their beliefs about their own social performance in less-biased ways than participants who completed only the EMA intervention, but this unbiased updating seems to be a mixed blessing. CBM-I seems to have mitigated the bias (which was observed in the EMA-only group) toward learning more from feedback on poor rather than good aspects of speech performance and having this type of unbiased updating is likely healthier for the CBM-I group (than the bias observed in the EMA-only group). However, CBM-I also seems to have mitigated the bias (that was observed in the EMA-only group) toward learning more from positive feedback than negative feedback. The EMA-only group’s bias toward learning from positive feedback would likely lead to more positive expectancies of one’s own social performance over time, which may be healthier than the CBM-I group’s unbiased learning from positive and negative feedback, although this remains to be more directly tested.

The effects observed over time regardless of the intervention condition suggest that tracking emotions and emotion regulation may also be helpful for social RL, possibly by increasing learning from positive feedback and decreasing learning from negative feedback. Given that all participants completed EMA, it is difficult to disentangle the mechanisms that might be at work. Emotion tracking might make individuals who are socially anxious more aware of the outcomes of their emotion regulation attempts, or it might encourage them to regulate in ways that promote positive affect, increasing their motivation to learn from positive social outcomes. Alternatively, they may pay more attention to when people treat them poorly and learn to make better decisions accordingly to avoid that treatment. Determining the specific mechanisms of change in social RL will be helpful in future studies.

Other studies in this data set have found that CBM-I decreased participants’ trait negative interpretation bias and increased their ability to use reappraisal day-to-day, but CBM-I was not associated with changes in cognitive flexibility and decreases in social anxiety symptoms occurred regardless of the intervention condition [37]. The fact that we found intervention-related changes in social RL and interpretation bias but nonspecific changes in symptoms raises interesting questions regarding the relationships among these cognitive, emotional, and behavioral processes. The current research design does not allow us to disentangle the temporal relations between changes in social RL and changes in interpretation bias and symptoms. However, it seems possible (and worth testing empirically in future research) that changes in social RL precede changes in social anxiety symptoms. For example, improvements in social decision-making accuracy may lead to more rewarding social interactions, which may decrease social avoidance and fear of negative evaluation.

There are several limitations worth noting in this study’s methodology. First, although the CBM-I intervention included mostly social scenarios, it also included several nonsocial scenarios. Its effects on social RL might have been stronger if the intervention specifically targeted negative interpretation bias in social situations. Second, this study might have been underpowered to detect the hypothesized 3-way interactions. Unfortunately, we did not conduct an a priori power analysis specific to these questions (as these data were part of a larger parent study). Relevant effect sizes in other studies with fewer participants were medium-large (eg, Cohen d=0.54 for participants with high social anxiety avoiding punishing faces more frequently than those with low social anxiety [13]). On the basis of a post hoc power analysis [60] using summary statistics from the study by Abraham and Hermann [13], our sample size of 106 was close to the recommended sample size of 111. However, sample sizes need to be larger to detect 3-way interactions versus the 2-way interactions tested in past studies [61]. As such, our sample may have been appropriately powered to detect 2-way interactions but may have been underpowered to detect higher-degree interactions. Third, some characteristics of our sample, including their relative homogeneity in age, race, and ethnicity and that they were not clinically diagnosed with social anxiety disorder, limit generalizability to clinical populations more broadly. Fourth, because we were only able to ethically collect data on social RL about one’s own social performance at the end of the study, we do not know if there were baseline differences in performance on this task between the CBM-I and EMA-only groups.

Given our initial findings of several positive changes in social RL associated with CBM-I, future research may help illuminate the mechanisms of change as well as how changes in social RL relate to changes in other important constructs, such as reduced anxious avoidance. Furthermore, studies may look at how social RL changes over the course of interventions other than CBM-I, such as cognitive behavioral or interpersonal therapy, may predict response to treatment. Future studies may also incorporate passive sensing to determine the optimal time and context to deliver CBM-I as a just-in-time adaptive intervention, reaching people when it can help them the most [62]. Updating rigidly negative expectancies and beliefs to effectively consider a range of feedback and information is key to healthy interpersonal and emotional functioning; continued research on social RL in social anxiety may help us understand why certain biases persist and how we might work to change them.