People’s decisions are influenced by their beliefs, which may be based on advice from other humans or, alternatively, on information from non-human sources such as road signs. But which sources do we find more reliable? Although scientists have studied the importance of social context in the way we process information, it is not fully understood how the brain processes information differently depending on who provides it.

Trudel et al. investigated the differences in the way humans evaluate information from human sources, such as advisors, compared to non-human sources, like inanimate objects, by monitoring brain activity and analyzing the results using a computational approach. In the experiments, 24 participants received a reward for locating a hidden dot on a circle under two different conditions: they either received a clue on the dot’s location from an image of a human face (social condition), representing an advisor, or from an inanimate object (non-social condition). Participants received information from many different advisors and inanimate objects, and the accuracy of the clues given by any of them varied from source to source. Each time, participants reported whether they thought the advice was reliable.

The results of monitoring the participants’ brain activity showed that they used different strategies when assessing the reliability of advice from a human than when the information came from a non-human source. Additionally, participants based their judgments about an advisor more strongly on past experiences with them, that is, if an advisor had given them good advice in the past, they were more likely to rely on their advice. Conversely, judgments about an inanimate object were based more strongly on recent experiences with that object.

Interestingly, participants were more certain when making judgments about the accuracy of cues given by advisors compared to inanimate objects, and they also updated their assessment of human sources less according to new evidence that contradicted their initial belief. This suggests that people may form more stable opinions about the reliability of sources when they receive information in social contexts, possibly because they expect more consistent behavior from humans. This stability in judgments about advisors was also reflected in the signal of brain areas that are often involved when interacting with others.

The work of Trudel et al. shows that even the suggestion that the source of a piece of advice is human can change how we process the information. This is especially important because humans spend increasingly more time in the digital world. Awareness of our biased assessment of human sources will have implications for designing interactive tools to guide human decision-making as well as strategies to develop critical thinking.

To behave adaptively, humans and many other species learn from their conspecifics which actions to take and which ones to avoid. The human ability to learn not only from direct interactions with the environment, but also from other people contributes critically to our behavioural repertoires as individuals and also to collective culture (Boyd and Richerson, 2009; Gweon, 2021). When seeking information about which courses of actions to pursue, we routinely turn to others for advice and adjust our own course of action based on the quality of their advice (Bang et al., 2017; Bang et al., 2020; Behrens et al., 2008; De Martino et al., 2017; Hertz et al., 2017). Clearly, it is important to learn over time whether an advice giver is predicting external events accurately. This process, sampling and learning from other people’s advice, is arguably a process that may be very similar to learning from non-social cues about the occurrence of external events (Akaishi et al., 2016; Trudel et al., 2021). In other words, the cognitive operations underlying this type of information search might occur both in social and in non-social scenarios. Therefore, by comparing information sampling from social versus non-social sources, we address a long-standing question in cognitive neuroscience, the degree to which any neural process is specialized for, or particularly linked to, social as opposed to non-social cognition (Chang and Dal Monte, 2018; Diaconescu et al., 2017; Frith and Frith, 2010; Frith and Frith, 2012; Grabenhorst and Schultz, 2021; Lockwood et al., 2020a; Lockwood et al., 2018; Soutschek et al., 2016; Wittmann et al., 2018; Rushworth et al., 2012).

Given their similarities, it is expected that both types of learning will depend on common neural mechanisms. However, given the importance and ubiquity of social learning, it may also be that the neural mechanisms that support learning from social advice are at least partially specialized and distinct from those concerned with learning that is guided by non-social sources. However, it is less clear on which level information is processed differently when it has a social or non-social origin. It has recently been argued that differences between social and non-social learning can be investigated on different levels of Marr’s information processing theory: differences could emerge at an input level (in terms of the stimuli that might drive social and non-social learning), at an algorithmic level or at a neural implementation level (Lockwood et al., 2020a). It might be that, at the algorithmic level, associative learning mechanisms are similar across social and non-social learning (Heyes, 2012). Other theories have argued that differences might emerge because goal-directed actions are attributed to social agents which allows for very different inferences to be made about hidden traits or beliefs (Shafto et al., 2012). Such inferences might fundamentally alter learning about social agents compared to non-social cues.

There is evidence that at least some parts of the brain are specialized for negotiating social situations (Noonan et al., 2014; Sallet et al., 2011; Sliwa and Freiwald, 2017). For example, neuroimaging studies in macaques show a positive correlation between group size and areas in the temporal lobe (Sallet et al., 2011), in an area that has similarities with the temporoparietal junction (TPJ) in humans (Mars et al., 2013). In turn, human TPJ and the macaque STS region have been associated and causally linked to the ability to infer other people’s beliefs (Hill et al., 2017; Schurz et al., 2017) and, in monkeys, to inferring the intended actions of others (Ong et al., 2021). Dorsomedial prefrontal cortex (dmPFC) is a second important node within this network again both in humans (Wittmann et al., 2018; Hampton et al., 2008; Nicolle et al., 2012; Piva et al., 2019; Suzuki et al., 2012) and in non-human primates (Noritake et al., 2018; Yoshida et al., 2012). For example, a recent study demonstrates dmPFC is causally implicated in social cognition and the maintenance of separate representations of oneself and of other individuals even when people interact (Wittmann et al., 2021). There are however, studies that report many common brain areas linked to reward and motivational processes in the context of both social and non-social behaviour suggesting at the very least that social information processing is not completely separate (Behrens et al., 2008; Boorman et al., 2013; Ruff and Fehr, 2014; Will et al., 2017; Zink et al., 2008).For example, tracking the probability of being correct or successful, often referred to as ‘confidence’ in one’s choice, has been shown to correlate with activation in perigenual anterior cingulate cortex in both social (Wittmann et al., 2016) and non-social (Bang and Fleming, 2018) settings. Here, we seek to understand the behavioural and neural computations that differentiate learning from social and non-social information sources and to further test at which level of information processing these differences occur.

It has been proven difficult to dissociate social as opposed to non-social cognitive functions. One reason is the lack of controlled within-subject experimental designs that have rarely been implemented in previous social neuroscience studies. Another reason might be the type of cognitive process that is studied; only cognitive functions that might plausibly occur in either social and non-social scenarios allow for a fair comparison. Here, we attempted to optimize our experimental design to address both these considerations in order to examine whether there are any fundamental differences in behavioural and neural computations when seeking information from social as opposed to non-social sources. First, we implemented a tightly controlled within-subject design with two experimental sessions that were matched in their statistical properties and only differed in their framing (social vs. non-social) and in their precise visual features. This ensured that differences between conditions were confined to the nature of the information source – social or non-social – and therefore that any differences in behaviour or neural activity could only attributed to the social/non-social difference. Second, we compared learning from other people’s advice with learning from non-social cues about the occurrence of external events. On every trial of the experiment, participants witnessed the precision with which a social cue (i.e., advice) or a non-social cue predicted a target. Over time, participants learned about the quality of these predictions and formed estimates of how confident they were in a cue, that is how precisely they expected the target to be predicted by the cue. We show that humans appear more certain in the performance accuracy of social predictors (advisors) compared to non-social predictors. This was evident in more stable confidence judgements across multiple timescales in the social as opposed to the non-social condition; there was a stronger reliance on representations made in the past and a weaker integration of new contradictory evidence. By using a computational approach, we could associate differences in the stability in confidence about social advisors to differences in the Bayesian estimates of participants’ subjective beliefs. We found that two brain areas, dmPFC and posterior TPJ (pTPJ) showed specificity in their processing of social information, not only in their average activation but also by encoding the identity of social cues in their multivariate pattern activation.

We used a social and non-social version of a previously validated paradigm (Trudel et al., 2021) during which participants learned about the performance of social and non-social cues in predicting a target on the circumference of a circle. On every trial, participants received advice on where to find the target (Figure 1 a, b , Figure 1—figure supplement 1). In the social version, faces represented advisors that predicted a target position (a flower’s position on the circumference of a circle) (Figure 1a). Some advisors predicted the target accurately and others less accurately. Participants were instructed that advisors had previously been players of a similar task during which time they could directly learn about the target location (Figure 1c, Figure 1—figure supplement 2). In other words, the other players had learned directly about the target location, while in the main experiment, participants could only infer the target locations from the predictions made by these previous players who now acted as advisors, but not directly from the target. Performance was defined by the size of an angular error, which represented the distance between the location of the predicted flower position (represented by a black dot) and the true position of the flower (represented by a yellow dot).

Figure 1 with 2 supplements see all

Download asset Open asset

In the non-social version, participants were instructed to collect fruits (the fruit position was now the target position and again it was represented by a yellow dot) that could either fall a small or large distance from the tree (this was now the predictor and again it was represented by a black dot). Importantly, task versions were completely matched in their experimental design and statistical properties inherent to the predictors and only differed in terms of their framing towards the participants. Thus, if there were differences in behaviour or neural activity between the conditions, then these could not have been attributed to any difference in protocol but only to a predisposition to treat the two types of predictors differently. Note that, the critical predictive stimuli and targets were, respectively, simply yellow and black dots appearing on a similar circle’s circumference in both conditions.

On every trial, participants indicated their confidence in the advisor’s performance. A symmetrical confidence interval (CI) with a random initial width appeared around the (social or non-social) prediction. The CI could be changed continuously to make it wider or narrower, by pressing buttons repeatedly (one button press resulted in a change of one step in the CI). In this way, participants provided what we refer to as an ‘interval setting’. Reward was received when the target fell within the chosen interval and the number of points increased when the interval size was small. A narrow interval could be set when participants selected a predictor with a small angular error and when they were certain about the predictor’s angular error. Hence, trial-wise confidence judgements allowed us to inspect participants’ beliefs about the performance of the advisor and their uncertainty in this performance estimate. In both social and non-social conditions, the participants’ beliefs about the predictor’s performance could be updated during the outcome phase when they witnessed the angular error between the predictive black dot and the target yellow dot. Participants were presented with multiple social and non-social predictors across six blocks. In both social and non-social settings, predictors varied in the accuracy with which they predicted the target, that is their average angular error (Figure 1c). For each predictor accuracy in the social condition, there was always a predictor in the non-social condition with a matched accuracy level.

Confidence in the performance of social and non-social advisors

We compared trial-by-trial interval setting in relation to the social and non-social advisors/predictors. When setting the interval, the participant’s aim was to minimize it while ensuring it still encompassed the final target position; points were won when it encompassed the target position but were greater when it was narrower. A given participant’s interval setting should, therefore, change in proportion to the participant’s expectations about the predictor’s angular error and their uncertainty about those expectations. Even though, on average, social and non-social sources did not differ in the precision with which they predicted the target (Figure 2—figure supplement 1), participants gave interval settings that differed in their relationships to the true performances of the social advisors compared to the non-social predictors. The interval setting was closer to the angular error in the social compared to the non-social sessions (Figure 2a, paired t-test: social vs. non-social, t(23) = −2.57, p = 0.017, 95% CI = [−0.36 −0.4]). Differences in interval setting might be due to generally lower performance in the non-social compared to social condition, or potentially due to fundamentally different learning processes utilized in either condition. We compared the mean reward amounts obtained by participants in the social and non-social conditions to determine whether there were overall performance differences. There was, however, no difference in the reward received by participants in the two conditions (mean reward: paired t-test social vs. non-social, t(23) = 0.8, p = 0.4, 95% CI = [−0.007 0.016]), suggesting that interval setting differences might not simply reflect better or worse performance but instead suggest quantitative differences in qualitatively similar learning mechanisms.

Figure 2 with 2 supplements see all

Download asset Open asset

Figure 2—source data 1

Source data include data shown in Figure 2a-c.

Further, it includes regression weights shown in Figure 2d for the effects of accuracy, and the effect of uncertainty. It also includes values for the learning index for the first and remaining updates, Figure 2e. Source data include values for both social and non-social conditions.

https://cdn.elifesciences.org/articles/71315/elife-71315-fig2-data1-v1.xlsx

Download elife-71315-fig2-data1-v1.xlsx

Estimating the angular error of an advisor accurately depends on how judgements are adjusted from trial to trial. We compared changes in interval settings on two timescales. First, we did this on a longer timescale, by comparing the absolute change between the current and next judgement of the same predictor (which might only be presented again several trials later, Figure 2b). Second, we did this on a within-trial timescale, by comparing changes of mind prior to the final judgement: participants adjusted the CI’s size by individual button presses, allowing inspections of changes of mind within a trial according to a specific sequence of button presses prior to the final judgement. Changes of mind within a trial were indicated by first becoming more confident (selecting smaller intervals) and then becoming less confident (selecting a bigger interval) before the final judgement (Figure 2c, inset). Participants made fewer adjustments of their confidence judgements across both timescales for social advisors: participants’ absolute changes in their confidence judgements from trial to trial were lower (Figure 2b, paired t-test: social vs. non-social, t(23) = −2.7, p = 0.0125, 95% CI = [−0.3 −0.04]) and they exhibited fewer changes of mind before committing to a final interval setting (paired t-test: social vs. non-social, t(23) = −2.4, p = 0.025, 95% CI = [−6.3 −0.45]). In sum, these results suggest that participants are more certain in their interval setting for social compared to non-social information sources, as they change estimates about an advisor less drastically across trials and commit to their judgement within each trial with less hesitation.

We tested the hypothesis that participants’ certainty in their own beliefs influenced their interval setting differently in social and non-social settings. We used a Bayesian model that allowed dissociation of two aspects of subjective beliefs: first, belief in how well the predictor will perform on a given trial (a belief in the predictor’s ‘accuracy’) and second, the participant’s uncertainty associated with the predictor’s predicted accuracy (‘uncertainty’ of the belief) (Figure 2d, Figure 2—figure supplement 2; Trudel et al., 2021). Both aspects of a belief might impact on the adjustment of the interval setting. Accuracy is the more likely belief parameter of the two to affect interval setting because accuracy reflects a point-estimate of the most likely angular error that an advisor has shown in the past. However, narrow interval settings, which have the potential to yield more reward if they encompass the target (Figure 2a) can only be employed when participants believe the advice given to be a true reflection of the target location (the advice is accurate) and when they are certain in their belief that the advice will be accurate (certainty in the advice’s accuracy). We, therefore, tested the impact of both belief parameters on trial-wise interval settings and examined whether their impact differed when making judgements about social or non-social cues. For both social and non-social predictors, participants were more confident and set a smaller interval when they believed the predictor to be accurate in their target prediction (paired t-test: social vs. non-social, t(23) = 0.6, p = 0.558, 95% CI = [−0.13 0.24]; one sample t-test: non-social, t(23) = 3.2, p = 0.003, 95% CI = [0.09 0.41]; one sample t-test: social, t(23) = 4, p < 0.0001, 95% CI = [0.15 0.46], Figure 2e, Figure 1—figure supplement 2). However, when judging non-social predictors, but not social predictors, interval settings were additionally widened as a function of subjective uncertainty: participants set a larger interval when they were more uncertain about the non-social cue’s accuracy to predict the target (paired t-test: social vs. non-social, t(23) = 3.5, p = 0.0018, 95% CI = [0.11 0.42]; one sample t-test: non-social, t(23) = −3.9, p < 0.0001, 95% CI = [−0.35 −0.1]; one sample t-test: social, t(23) = 0.6, p = 0.56, 95% CI = [−0.1 0.19], Figure 2e). The difference between social and non-social conditions was replicated when using a mixed-effects model (Figure 2—figure supplement 1). Uncertainty about the accuracy of social advisors, however, did not have the same impact: participants’ interval settings for the social advisors appeared only to be guided by the peak of the Bayesian probability distribution, as indicated by the accuracy effect.

We have seen that participants have less changeable estimates of the performances of social advisors compared to non-social advisors (Figure 2b), and this relative increase in certainty about the value of social advice as opposed to non-social information is also reflected in the Bayesian estimates of participants’ subjective beliefs (Figure 2d, e). One possibility is that this increased certainty is caused by relying on a longer-term memory of the observed advice in the social compared to the non-social context. Participants might form an opinion about the performance of social advisors more quickly and are therefore less likely to change it in face of new and possibly contradictory evidence. We therefore tested whether the degree of error-driven behavioural adaptation is different between social and non-social conditions. In other words, we tested whether participants changed their interval setting from one trial to the next to a smaller degree as a function of the discrepancy between their interval setting and the angular error (i.e., prediction error). We calculated a trial-by-trial learning index, similar to a learning rate in a reinforcement learning (RL) model (Barto, 1998; Wittmann et al., 2020; Lockwood and Klein-Flügge, 2020b), that measured the change between the current and next interval setting given the prediction error (Figure 2f; Methods, Equation 2 and 3). A smaller learning index shows less error-driven adjustment of one’s interval setting. Learning indices between conditions were similar for the first observation but declined more steeply for social compared to non-social predictors for the remaining updates (repeated-measures analysis of variance [ANOVA]: interaction between group (social, non-social) × time point (first update, other remaining updates), F(1,23) = 5.8, p = 0.024, Figure 2f). Hence, after the first update, interval setting in the social condition was subsequently less influenced by prediction errors that was the case in the non-social case (paired t-test: social vs. non-social for remaining only, t(23) = 2.7, p = 0.01, 95% CI = [−0.95 −0.13], Figure 2f). As a consequence, judgements relied more on observations made further in the past in the social condition. This is consistent with the observation that participants repeatedly set similar intervals across time in the social condition (i.e., fewer changes across trials) (Figure 2b), and that these settings were less impacted by their own subjective uncertainty (Figure 2e).

The impact of noise in belief updating in social and non-social conditions

So far, we have shown that, in comparison to non-social predictors, participants changed their interval settings about social advisors less drastically across time, relied on observations made further in the past, and were less impacted by their subjective uncertainty when they did so (Figure 2). In further exploratory analyses, we used Bayesian simulation analyses to investigate whether a common mechanism might underlie these behavioural differences. We tested whether the integration of new evidence differed between social and non-social conditions; for example, recent observations might be weighted more strongly for non-social cues while past observations might be weighted more strongly for social cues. Similar ideas were tested in previous studies, when comparing the learning rate (i.e., the speed of learning) in environments of different volatilities (Behrens et al., 2007; Meder et al., 2017). We tested these ideas using established ways of changing the speed of learning during Bayesian updates (Allenmark et al., 2018; Yu and Cohen, 2008). We hypothesized that participants reduce their uncertainty quicker when observing social information. Vice versa, we hypothesized a less steep decline of uncertainty when observing non-social information, indicating that new information can be flexibly integrated during the belief update (Figure 3a).

Figure 3 with 1 supplement see all

Download asset Open asset

We manipulated the amount of uncertainty in the Bayesian model by adding a uniform distribution to each belief update (Figure 3b, c; Equations 10 and 11). Consequently, the distribution’s width increases and is more strongly impacted by recent observations (see example in Figure 3—figure supplement 1). We used these modified Bayesian models to simulate trial-wise interval setting for each participant according to the observations they made by selecting a particular advisor in the social condition or other predictor in the non-social condition. We simulated CIs at each trial. We then used these to examine whether an increase in noise led to simulation behaviour that resembled behavioural patterns observed in non-social conditions that were different to behavioural patterns observed in the social condition.

First, we repeated the linear regression analysis and hypothesized that interval settings that were simulated from noisy Bayesian models would also show a greater negative ‘predictor uncertainty’ effect on interval setting resembling the effect we had observed in the non-social condition (Figure 2e). This was indeed the case when using the noisy Bayesian model: irrespective of social or non-social condition, the addition of noise (increasing weight of the uniform distribution to each belief update) led to an increasingly negative ‘predictor uncertainty’ effect on confidence judgement (new Figure 3d). The absence of difference between the social and non-social conditions in the simulations suggests that an increase in the Bayesian noise is sufficient to induce a negative impact of ‘predictor uncertainty’ on interval setting. Hence, we can conclude that different degrees of noise in the updating process are sufficient to cause differences observed between social and non-social conditions. Next, we used these simulated interval settings and repeated the descriptive behavioural analyses (Figure 2b, f). An increase in noise led to greater changes of confidence across time and a higher learning index (Figure 3—figure supplement 1). In summary, the Bayesian simulations offer a conceptual explanation that can account for both the differences in learning and the difference in uncertainty processing that exist between social and non-social conditions. The key insight conveyed by the Bayesian simulations is that a wider, more uncertain belief distribution changes more quickly. Correspondingly, in the non-social condition, participants express more uncertainty in their confidence estimate when they set the interval, and they also change their beliefs more quickly. Therefore, noisy Bayesian updating can account for key differences between social and non-social condition.

dmPFC and pTPJ covary with social confidence and encode social advisor identity

Next, we tested whether there are brain areas that covary with interval setting that are made about social compared to non-social advisors. Behaviourally, we have shown that judgements about social advisors might reflect the presence and influence of relatively constant and unchanging represen tations in memory. Therefore, a combination of the trial-wise interval setting made in the current (time point t) and past trial (time point t − 1) for the same predictor was used to test for differences between social and non-social advisor judgements (contrast: current + past interval setting). We sign-reversed the interval setting measure such that a higher index now reflected greater confidence in the selected predictor. We focussed on two a priori independent regions of interest (ROIs), dmPFC (Wittmann et al., 2016) and pTPJ (Mars et al., 2013) because of their previous association with inference about beliefs held by others (Schurz et al., 2017; Hampton et al., 2008; Nicolle et al., 2012; Piva et al., 2019; Suzuki et al., 2012; Wittmann et al., 2016; Saxe, 2006); a process that is crucial to the current task. Hence, we tested whether blood-oxygen-level-dependent (BOLD) signal in these brain areas changed in tandem with variation in the intervals participants set. This parametric contrast thus captures trial-by-trial variation in the confidence that participants have in the predictor and is independent of the mean confidence levels. Note, however, that the mean confidence level was not significantly different between social advisors and non-social predictors (paired t-test: mean confidence, social vs. non-social, t(23) = −1.7, p = 0.11, 95% CI = [−0.66 0.07]), even though we observed greater stability of interval setting for social advisors (Figure 2b). Activation in both dmPFC and pTPJ covaried with combined confidence judgements significantly more strongly for social compared to non-social advisors (Figure 4a, paired t-test: social vs. non-social for dmPFC, t(23) = −2.5, p = 0.019, 95% CI = [3 31]; Figure 4b, paired t-test: social vs. non-social for pTPJ: t(23) = −2.3, p = 0.03, 95% CI = [1 22.5]). Next, we performed an exploratory whole-brain analysis and found that these activations in dmPFC and pTPJ were also detectable when examining activity across the brain and using standard cluster-correction techniques to adjust for multiple comparisons when making judgements about social advisors (family-wise error (FWE) cluster corrected with z-score >2.3 and p < 0.05, Supplementary file 1). In addition, activity was significantly related to the same measure in the social condition after whole-brain cluster-correction, in some other areas previously linked with social cognition such as the medial frontal pole and posterior cingulate/medial parietal cortex (Figure 4c, d, for a complete list of cluster-corrected brain areas, see Supplementary file 1).

Figure 4 with 4 supplements see all

Download asset Open asset

Figure 4—source data 1

Source data include data shown in Figure 4a (contrast: current + past interval setting in dorsomedial prefrontal cortex [dmPFC] region of interest [ROI]), Figure 4b (contrast: current + past interval setting in posterior temporoparietal junction [pTPJ] ROI), Figure 4e (separate effects of current and past interval setting in dmPFC ROI for social condition only), Figure 4e (separate effects of current and past interval setting in pTPJ ROI for social condition only), Figure 4h (Exemplar Discriminability Index [EDI] values for dmPFC ROI), and Figure 4h (EDI values for pTPJ ROI).

Source data include values for both social and non-social conditions.

https://cdn.elifesciences.org/articles/71315/elife-71315-fig4-data1-v1.xlsx

Download elife-71315-fig4-data1-v1.xlsx

Our behavioural results showed that social judgements might be especially impacted by beliefs formed over longer timescales: for example, judgements about social advisors changed less drastically across time, even if new information at the time of the trial outcome was contradictory to one’s current belief (Figure 2f). We tested whether such a differentiation between current (t) and past (t − 1) interval setting was also apparent in the neural activation profile of dmPFC and pTPJ. In other words, we tested whether pTPJ and dmPFC displayed an activation profile consistent with a greater influence of interval settings over the longer term, as opposed to on just the current trial, to a greater degree in the social as opposed to the non-social condition. Consistent with the behavioural results, activation within both areas covaried with interval settings made over longer timescales rather than just on the current trial (one sample t-test, social: dmPFC, past vs. current: t(23) = 3.04, p = 0.0058, 95% CI = [7.6 39.8]; pTPJ, past vs. current, t(23) = 2.25, p = 0.034, 95% CI = [1.1 26.6]; Figure 3e, f; Figure 4—figure supplement 1). There was no difference in the strength of neural activity related to interval setting at the two time scales in the non-social condition (Figure 4—figure supplement 2).

So far, we have focussed on how the univariate activity in dmPFC and pTPJ covaried with the participants’ estimates of the accuracy of predictors. Next, we examined whether the same brain regions carried additional information about the predictors that is orthogonal to what we have examined so far; we sought evidence as to whether the two areas encoded the identities of the cues. We applied a representational similarity analysis (RSA) to BOLD activity patterns evoked by each face and each fruit presentation at the time of response during the confidence phase. We compared the pattern dissimilarity across voxels (measured by the Euclidean distance) in the same a priori independent ROIs in dmPFC and pTPJ as examined previously. For each condition, the Exemplar Discriminability Index (EDI) was calculated which reflects the average dissimilarity in activity elicited across presentations of the same cue (Figure 3g, orange area) compared to the average dissimilarity across presentations of different cues (Figure 3g, red area) (Bang et al., 2020). Hence, a negative EDI indicates that an area’s activity changes less when the same advisor/predictor is presented again than when different advisors/predictors are presented suggesting that it encodes representations of the identity of predictors/advisors. The patterns of activation change in both dmPFC and pTPJ suggested that both areas encoded the identities of social cues more strongly than non-social cues (Figure 3h; repeated-measures ANOVA: main effect of condition (social, non-social) in 2 (condition: social, non-social) × 2 (area: pTPJ, dmPFC) ANOVA: F(1,21) = 5, p = 0.036). An exploratory whole-brain searchlight analysis was performed to test the specificity of dmPFC and pTPJ encoding the identity of social compared to non-social cues. No other cluster-corrected effects were found (Figure 4—figure supplement 3). Additional control analyses show that neural differences between social and non-social conditions are not due to the visually different set of stimuli used in the experiment but are instead representing fundamental differences in processing social compared to non-social information (Figure 4—figure supplement 4). In summary, the conjunction, univariate and multivariate analyses demonstrate that dmPFC and pTPJ represent beliefs about social advisors that develop over a longer timescale and encode the identities of the social advisors.

We examined the behavioural and neural computations that differentiate learning from social and non-social information sources. Using a tightly controlled within-subject design in which conditions only differed in their social or non-social framing, we asked participants to make trial-wise confidence judgements, by setting a visual interval, about the advice of social and non-social information sources. Ideally, confidence should increase with the accuracy of one’s choice (Bang and Fleming, 2018; Kiani and Shadlen, 2009; Lak et al., 2020). Here, participants did not match their confidence to the likely accuracy of their own performance, but instead to the performance of another social or non-social advisor. Participants used different strategies when setting intervals to express their confidence in the performances of social advisors as opposed to non-social advisors. A possible explanation might be that participants have a better insight into the abilities of social cues – typically other agents – than non-social cues – typically inanimate objects. We used a computational approach to test the interplay of two belief estimates that impact interval setting: the belief in how accurate the advice/prediction is and the subjective uncertainty in that belief. Intervals were impacted by several factors; however, it was the influence of subjective uncertainty about the performance estimates that distinguished social and non-social interval setting. Participants selected a greater interval size when making their confidence judgements on non-social trials, indicating that they were more uncertain about the performance accuracies of the non-social cues compared to the social cues.

Being less influenced by ones’ uncertainty about the value of social advice might help explain why people sometimes fail to use evidence to update their assessments of social advisors (Fleming et al., 2018; Rollwage et al., 2020). Detecting uncertainty in environmental contingencies or about one’s own beliefs is achieved with fine-tuned metacognitive abilities (Miyamoto et al., 2021) and is essential for behavioural adaptation (Trudel et al., 2021; Badre et al., 2012). Participants were less influenced by their subjective uncertainty when estimating the performances of social advisors. Moreover, they changed and updated their beliefs about social cues less than about non-social cues when they observed the outcomes of their predictions. Repeating a particular interval size instead of adjusting it from trial to trial might be a better strategy in the current experiment because all advisors had, on average, a stable performance level. However, opting for such a strategy only in the context of social advisors might reflect a natural tendency to expect other social agents to behave consistently and exhibit stable performance levels as soon as an initial learning period is passed. It may be because we make just such assumptions that past observations are used to predict performance levels that people are likely to exhibit next (Wittmann et al., 2021; Wittmann et al., 2016). An alternative explanation might be that participants experience a steeper decline of subjective uncertainty in their beliefs about the accuracy of social advice, resulting in a narrower prior distribution, during the next encounter with the same advisor. During a series of exploratory analyses, we used Bayesian simulations to investigate how uncertainty about beliefs changed from trial to trial and showed that belief updates about non-social cues were consistent with a noisier update process that diminished the impact of experiences over the longer term. From a Bayesian perspective, greater certainty about the value of advice means that contradictory evidence will need to be stronger to alter one’s beliefs. In the absence of such evidence, a Bayesian agent is more likely to repeat previous judgements. Just as in a confirmation bias (Kappes et al., 2020), such a perspective suggests that once we are more certain about others’ features, for example, their character traits, we are less likely to change our opinions about them.

One potential explanation for the assumption of stable performance for social but not non-social predictors might be that participants attribute intentions and motivations to social agents. Even if the social and non-social evidence are the same, the belief that a social actor might have a goal may affect the inferences made from the same piece of information (Shafto et al., 2012). Social advisors first learnt about the target’s distribution and accordingly gave advice on where to find the target. If the social agents are credited with goal-directed behaviour then it might be assumed that the goals remain relatively constant; this might lead participants to assume stability in the performances of social advisors. However, such goal-directed intentions might not be attributed to non-social cues, thereby making judgements inherently more uncertain and changeable across time. Such an account, focussing on differences in attribution in social settings aligns with a recent suggestion that any attempt to identify similarities or differences between social and non-social processes can occur at any one of a number of the levels in Marr’s information theory (Lockwood et al., 2020a). Here, we found that the same algorithm was able to explain social and non-social learning (a qualitatively similar computational model could explain both). However, the extent to which the algorithm was recruited when learning about social compared to non-social information differed. We observed a greater impact of uncertainty on judgements about social compared to non-social information. We have shown evidence for a degree of specialization when assessing social advisors as opposed to non-social cues. At the neural level we focussed on two brain areas, dmPFC and pTP. These brain areas carry signals associated with belief inferences about others, for example how confident we think other people are about their own choices (Bang et al., 2022). Further, recent combined fMRI–TMS studies have also demonstrated the causal importance of these activity patterns for social inference (Hill et al., 2017; Wittmann et al., 2021). Similar processes might be at play in the current study during which participants make inferences about the performance of the social advisor in locating the target. We examined dmPFC and pTPJ, using ROIs derived from previous studies, and showed that these areas represent a combination of present and past confidence judgements that are made about social advisors and that they do this more strongly for social than for non-social cues.

A long-standing question in cognitive neuroscience is the degree to which any neural process is specialized for, or particularly linked to, social as opposed to non-social cognition. The current results suggest that dmPFC and pTPJ activity patterns are different during social cognition and that they carry information that is related to the performance differences that are seen at a behavioural level during a social as opposed to non-social task. Interestingly, it seems that activation in dmPFC and pTPJ was particularly prominent when representing judgements from previous encounters compared to the present judgement. This was apparent not only in univariate activity in these areas that covaried with confidence in advice but also in the multivariate activity patterns that encoded the identity of social agents; RSA showed that activation patterns in both areas encoded the identity of social cues better than the identity of non-social cues. The absence of similar differences in visual cortex suggests that the effects cannot be attributed to basic visual differences between the stimuli used in the two tasks. Hence, pTPJ and dmPFC seem particularly linked to the navigation of social interactions.

We have deposited the choice raw data used for the analyses in the OSF repository at https://osf.io/wjuze/. We also provide fMRI beta coefficients for the contrasts shown in Figure 4. Source data is provided with this paper. The above OSF repository includes code that recreates figures depicted in the main text. For the full Bayesian pipeline, which was used to derive regressors for the behavioural GLM depicted in Figure 2e, please refer to the following OSF repository: https://doi.org/10.17605/OSF.IO/D5QZW.

1. 1. Trudel N
   2. Lockwood P
   3. Rushworth M
   4. Wittmann M

   (2023) Open Science Framework

   Neural activity tracking identity and confidence in social information.

   https://doi.org/10.17605/OSF.IO/D5QZW

1. 1. Akaishi R
   2. Kolling N
   3. Brown JW
   4. Rushworth M

   (2016) Neural mechanisms of credit assignment in a multicue environment

   The Journal of Neuroscience 36:1096–1112.

   https://doi.org/10.1523/JNEUROSCI.3159-15.2016

   • PubMed
   • Google Scholar
2. 1. Allenmark F
   2. Müller HJ
   3. Shi Z

   (2018) Inter-Trial effects in visual pop-out search: factorial comparison of Bayesian updating models

   PLOS Computational Biology 14:e1006328.

   https://doi.org/10.1371/journal.pcbi.1006328

   • PubMed
   • Google Scholar
3. 1. Apps MAJ
   2. Rushworth MFS
   3. Chang SWC

   (2016) The anterior cingulate gyrus and social cognition: tracking the motivation of others

   Neuron 90:692–707.

   https://doi.org/10.1016/j.neuron.2016.04.018

   • PubMed
   • Google Scholar
4. 1. Apps MAJ
   2. Sallet J

   (2017) Social learning in the medial prefrontal cortex

   Trends in Cognitive Sciences 21:151–152.

   https://doi.org/10.1016/j.tics.2017.01.008

   • PubMed
   • Google Scholar
5. 1. Badre D
   2. Doll BB
   3. Long NM
   4. Frank MJ

   (2012) Rostrolateral prefrontal cortex and individual differences in uncertainty-driven exploration

   Neuron 73:595–607.

   https://doi.org/10.1016/j.neuron.2011.12.025

   • PubMed
   • Google Scholar
6. 1. Bang D
   2. Aitchison L
   3. Moran R
   4. Herce Castanon S
   5. Rafiee B
   6. Mahmoodi A
   7. Lau JYF
   8. Latham PE
   9. Bahrami B
   10. Summerfield C

   (2017) Confidence matching in group decision-making

   Nature Human Behaviour 1:0117.

   https://doi.org/10.1038/s41562-017-0117

   • Google Scholar
7. 1. Bang D
   2. Fleming SM

   (2018) Distinct encoding of decision confidence in human medial prefrontal cortex

   PNAS 115:6082–6087.

   https://doi.org/10.1073/pnas.1800795115

   • PubMed
   • Google Scholar
8. 1. Bang D
   2. Ershadmanesh S
   3. Nili H
   4. Fleming SM

   (2020) Private-public mappings in human prefrontal cortex

   eLife 9:e56477.

   https://doi.org/10.7554/eLife.56477

   • PubMed
   • Google Scholar
9. 1. Bang D
   2. Moran R
   3. Daw ND
   4. Fleming SM

   (2022) Neurocomputational mechanisms of confidence in self and others

   Nature Communications 13:4238.

   https://doi.org/10.1038/s41467-022-31674-w

   • PubMed
   • Google Scholar
10. 1. Barto AG

    (1998) Reinforcement learning

    IFAC Proceedings Volumes 31:5.

    https://doi.org/10.1016/S1474-6670(17)38315-5

    • Google Scholar
11. 1. Behrens TEJ
    2. Woolrich MW
    3. Walton ME
    4. Rushworth MFS

    (2007) Learning the value of information in an uncertain world

    Nature Neuroscience 10:1214–1221.

    https://doi.org/10.1038/nn1954

    • PubMed
    • Google Scholar
12. 1. Behrens TEJ
    2. Hunt LT
    3. Woolrich MW
    4. Rushworth MFS

    (2008) Associative learning of social value

    Nature 456:245–249.

    https://doi.org/10.1038/nature07538

    • PubMed
    • Google Scholar
13. 1. Boorman ED
    2. Behrens TE
    3. Rushworth MF

    (2011) Counterfactual choice and learning in a neural network centered on human lateral frontopolar cortex

    PLOS Biology 9:e1001093.

    https://doi.org/10.1371/journal.pbio.1001093

    • PubMed
    • Google Scholar
14. 1. Boorman ED
    2. O’Doherty JP
    3. Adolphs R
    4. Rangel A

    (2013) The behavioral and neural mechanisms underlying the tracking of expertise

    Neuron 80:1558–1571.

    https://doi.org/10.1016/j.neuron.2013.10.024

    • PubMed
    • Google Scholar
15. 1. Boyd R
    2. Richerson PJ

    (2009) Culture and the evolution of human cooperation

    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 364:3281–3288.

    https://doi.org/10.1098/rstb.2009.0134

    • PubMed
    • Google Scholar
16. 1. Caruana F
    2. Gerbella M
    3. Avanzini P
    4. Gozzo F
    5. Pelliccia V
    6. Mai R
    7. Abdollahi RO
    8. Cardinale F
    9. Sartori I
    10. Lo Russo G
    11. Rizzolatti G

    (2018) Motor and emotional behaviours elicited by electrical stimulation of the human cingulate cortex

    Brain 141:3035–3051.

    https://doi.org/10.1093/brain/awy219

    • PubMed
    • Google Scholar
17. 1. Chang SWC
    2. Dal Monte O

    (2018) Shining light on social learning circuits

    Trends in Cognitive Sciences 22:673–675.

    https://doi.org/10.1016/j.tics.2018.05.002

    • PubMed
    • Google Scholar
18. 1. De Martino B
    2. Bobadilla-Suarez S
    3. Nouguchi T
    4. Sharot T
    5. Love BC

    (2017) Social information is integrated into value and confidence judgments according to its reliability

    The Journal of Neuroscience 37:6066–6074.

    https://doi.org/10.1523/JNEUROSCI.3880-16.2017

    • PubMed
    • Google Scholar
19. 1. Diaconescu AO
    2. Mathys C
    3. Weber LAE
    4. Kasper L
    5. Mauer J
    6. Stephan KE

    (2017) Hierarchical prediction errors in midbrain and septum during social learning

    Social Cognitive and Affective Neuroscience 12:618–634.

    https://doi.org/10.1093/scan/nsw171

    • PubMed
    • Google Scholar
20. 1. Fleming SM
    2. van der Putten EJ
    3. Daw ND

    (2018) Neural mediators of changes of mind about perceptual decisions

    Nature Neuroscience 21:617–624.

    https://doi.org/10.1038/s41593-018-0104-6

    • PubMed
    • Google Scholar
21. 1. Frith C
    2. Frith U

    (2010) Learning from others: introduction to the special review series on social neuroscience

    Neuron 65:739–743.

    https://doi.org/10.1016/j.neuron.2010.03.015

    • PubMed
    • Google Scholar
22. 1. Frith CD
    2. Frith U

    (2012) Mechanisms of social cognition

    Annual Review of Psychology 63:287–313.

    https://doi.org/10.1146/annurev-psych-120710-100449

    • PubMed
    • Google Scholar
23. 1. Grabenhorst F
    2. Schultz W

    (2021) Functions of primate amygdala neurons in economic decisions and social decision simulation

    Behavioural Brain Research 409:113318.

    https://doi.org/10.1016/j.bbr.2021.113318

    • PubMed
    • Google Scholar
24. 1. Gweon H

    (2021) Inferential social learning: cognitive foundations of human social learning and teaching

    Trends in Cognitive Sciences 25:896–910.

    https://doi.org/10.1016/j.tics.2021.07.008

    • PubMed
    • Google Scholar
25. 1. Hampton AN
    2. Bossaerts P
    3. O’Doherty JP

    (2008) Neural correlates of mentalizing-related computations during strategic interactions in humans

    PNAS 105:6741–6746.

    https://doi.org/10.1073/pnas.0711099105

    • PubMed
    • Google Scholar
26. 1. Hertz U
    2. Palminteri S
    3. Brunetti S
    4. Olesen C
    5. Frith CD
    6. Bahrami B

    (2017) Neural computations underpinning the strategic management of influence in advice giving

    Nature Communications 8:2191.

    https://doi.org/10.1038/s41467-017-02314-5

    • PubMed
    • Google Scholar
27. 1. Heyes C

    (2012) What’s social about social learning?

    Journal of Comparative Psychology 126:193–202.

    https://doi.org/10.1037/a0025180

    • PubMed
    • Google Scholar
28. 1. Hill CA
    2. Suzuki S
    3. Polania R
    4. Moisa M
    5. O’Doherty JP
    6. Ruff CC

    (2017) A causal account of the brain network computations underlying strategic social behavior

    Nature Neuroscience 20:1142–1149.

    https://doi.org/10.1038/nn.4602

    • PubMed
    • Google Scholar
29. 1. Kappes A
    2. Harvey AH
    3. Lohrenz T
    4. Montague PR
    5. Sharot T

    (2020) Confirmation bias in the utilization of others’ opinion strength

    Nature Neuroscience 23:130–137.

    https://doi.org/10.1038/s41593-019-0549-2

    • PubMed
    • Google Scholar
30. 1. Kiani R
    2. Shadlen MN

    (2009) Representation of confidence associated with a decision by neurons in the parietal cortex

    Science 324:759–764.

    https://doi.org/10.1126/science.1169405

    • PubMed
    • Google Scholar
31. 1. Lak A
    2. Okun M
    3. Moss MM
    4. Gurnani H
    5. Farrell K
    6. Wells MJ
    7. Reddy CB
    8. Kepecs A
    9. Harris KD
    10. Carandini M

    (2020) Dopaminergic and prefrontal basis of learning from sensory confidence and reward value

    Neuron 105:700–711.

    https://doi.org/10.1016/j.neuron.2019.11.018

    • PubMed
    • Google Scholar
32. 1. Lockwood PL
    2. Wittmann MK
    3. Apps MAJ
    4. Klein-Flügge MC
    5. Crockett MJ
    6. Humphreys GW
    7. Rushworth MFS

    (2018) Neural mechanisms for learning self and other ownership

    Nature Communications 9:4747.

    https://doi.org/10.1038/s41467-018-07231-9

    • PubMed
    • Google Scholar
33. 1. Lockwood PL
    2. Apps MAJ
    3. Chang SWC

    (2020a) Is there a “ social” brain? implementations and algorithms

    Trends in Cognitive Sciences 24:802–813.

    https://doi.org/10.1016/j.tics.2020.06.011

    • PubMed
    • Google Scholar
34. 1. Lockwood PL
    2. Klein-Flügge MC

    (2020b) Computational modelling of social cognition and behaviour-a reinforcement learning primer

    Social Cognitive and Affective Neuroscience 16:761–771.

    https://doi.org/10.1093/scan/nsaa040

    • PubMed
    • Google Scholar
35. 1. Lockwood PL
    2. Wittmann MK
    3. Nili H
    4. Matsumoto-Ryan M
    5. Abdurahman A
    6. Cutler J
    7. Husain M
    8. Apps MAJ

    (2022) Distinct neural representations for prosocial and self-benefiting effort

    Current Biology 32:4172–4185.

    https://doi.org/10.1016/j.cub.2022.08.010

    • PubMed
    • Google Scholar
36. 1. Mars RB
    2. Jbabdi S
    3. Sallet J
    4. O’Reilly JX
    5. Croxson PL
    6. Olivier E
    7. Noonan MP
    8. Bergmann C
    9. Mitchell AS
    10. Baxter MG
    11. Behrens TEJ
    12. Johansen-Berg H
    13. Tomassini V
    14. Miller KL
    15. Rushworth MFS

    (2011) Diffusion-Weighted imaging tractography-based parcellation of the human parietal cortex and comparison with human and macaque resting-state functional connectivity

    The Journal of Neuroscience 31:4087–4100.

    https://doi.org/10.1523/JNEUROSCI.5102-10.2011

    • PubMed
    • Google Scholar
37. 1. Mars RB
    2. Sallet J
    3. Neubert FX
    4. Rushworth MFS

    (2013) Connectivity profiles reveal the relationship between brain areas for social cognition in human and monkey temporoparietal cortex

    PNAS 110:10806–10811.

    https://doi.org/10.1073/pnas.1302956110

    • PubMed
    • Google Scholar
38. 1. Meder D
    2. Kolling N
    3. Verhagen L
    4. Wittmann MK
    5. Scholl J
    6. Madsen KH
    7. Hulme OJ
    8. Behrens TEJ
    9. Rushworth MFS

    (2017) Simultaneous representation of a spectrum of dynamically changing value estimates during decision making

    Nature Communications 8:1942.

    https://doi.org/10.1038/s41467-017-02169-w

    • PubMed
    • Google Scholar
39. 1. Miyamoto K
    2. Trudel N
    3. Kamermans K
    4. Lim MC
    5. Lazari A
    6. Verhagen L
    7. Wittmann MK
    8. Rushworth MFS

    (2021) Identification and disruption of a neural mechanism for accumulating prospective metacognitive information prior to decision-making

    Neuron 109:1396–1408.

    https://doi.org/10.1016/j.neuron.2021.02.024

    • PubMed
    • Google Scholar
40. 1. Nicolle A
    2. Klein-Flügge MC
    3. Hunt LT
    4. Vlaev I
    5. Dolan RJ
    6. Behrens TEJ

    (2012) An agent independent axis for executed and modeled choice in medial prefrontal cortex

    Neuron 75:1114–1121.

    https://doi.org/10.1016/j.neuron.2012.07.023

    • PubMed
    • Google Scholar
41. 1. Nili H
    2. Wingfield C
    3. Walther A
    4. Su L
    5. Marslen-Wilson W
    6. Kriegeskorte N

    (2014) A toolbox for representational similarity analysis

    PLOS Computational Biology 10:e1003553.

    https://doi.org/10.1371/journal.pcbi.1003553

    • PubMed
    • Google Scholar
42. 1. Noonan MP
    2. Sallet J
    3. Mars RB
    4. Neubert FX
    5. O’Reilly JX
    6. Andersson JL
    7. Mitchell AS
    8. Bell AH
    9. Miller KL
    10. Rushworth MFS

    (2014) A neural circuit covarying with social hierarchy in macaques

    PLOS Biology 12:e1001940.

    https://doi.org/10.1371/journal.pbio.1001940

    • PubMed
    • Google Scholar
43. 1. Noritake A
    2. Ninomiya T
    3. Isoda M

    (2018) Social reward monitoring and valuation in the macaque brain

    Nature Neuroscience 21:1452–1462.

    https://doi.org/10.1038/s41593-018-0229-7

    • PubMed
    • Google Scholar
44. 1. Ong WS
    2. Madlon-Kay S
    3. Platt ML

    (2021) Neuronal correlates of strategic cooperation in monkeys

    Nature Neuroscience 24:116–128.

    https://doi.org/10.1038/s41593-020-00746-9

    • PubMed
    • Google Scholar
45. 1. Piva M
    2. Velnoskey K
    3. Jia R
    4. Nair A
    5. Levy I
    6. Chang SWC

    (2019) The dorsomedial prefrontal cortex computes task-invariant relative subjective value for self and other

    eLife 8:e44939.

    https://doi.org/10.7554/eLife.44939

    • PubMed
    • Google Scholar
46. 1. Rollwage M
    2. Loosen A
    3. Hauser TU
    4. Moran R
    5. Dolan RJ
    6. Fleming SM

    (2020) Confidence drives a neural confirmation bias

    Nature Communications 11:2634.

    https://doi.org/10.1038/s41467-020-16278-6

    • PubMed
    • Google Scholar
47. 1. Ruff CC
    2. Fehr E

    (2014) The neurobiology of rewards and values in social decision making

    Nature Reviews. Neuroscience 15:549–562.

    https://doi.org/10.1038/nrn3776

    • PubMed
    • Google Scholar
48. 1. Rushworth MFS
    2. Kolling N
    3. Sallet J
    4. Mars RB

    (2012) Valuation and decision-making in frontal cortex: one or many serial or parallel systems?

    Current Opinion in Neurobiology 22:946–955.

    https://doi.org/10.1016/j.conb.2012.04.011

    • PubMed
    • Google Scholar
49. 1. Sallet J
    2. Mars RB
    3. Noonan MP
    4. Andersson JL
    5. O’Reilly JX
    6. Jbabdi S
    7. Croxson PL
    8. Jenkinson M
    9. Miller KL
    10. Rushworth MFS

    (2011) Social network size affects neural circuits in macaques

    Science 334:697–700.

    https://doi.org/10.1126/science.1210027

    • PubMed
    • Google Scholar
50. 1. Saxe R

    (2006) Uniquely human social cognition

    Current Opinion in Neurobiology 16:235–239.

    https://doi.org/10.1016/j.conb.2006.03.001

    • PubMed
    • Google Scholar
51. 1. Schurz M
    2. Tholen MG
    3. Perner J
    4. Mars RB
    5. Sallet J

    (2017) Specifying the brain anatomy underlying temporo-parietal junction activations for theory of mind: a review using probabilistic atlases from different imaging modalities

    Human Brain Mapping 38:4788–4805.

    https://doi.org/10.1002/hbm.23675

    • PubMed
    • Google Scholar
52. 1. Shafto P
    2. Goodman ND
    3. Frank MC

    (2012) Learning from others: the consequences of psychological Reasoning for human learning

    Perspectives on Psychological Science 7:341–351.

    https://doi.org/10.1177/1745691612448481

    • PubMed
    • Google Scholar
53. 1. Sliwa J
    2. Freiwald WA

    (2017) A dedicated network for social interaction processing in the primate brain

    Science 356:745–749.

    https://doi.org/10.1126/science.aam6383

    • PubMed
    • Google Scholar
54. 1. Smith SM
    2. Jenkinson M
    3. Woolrich MW
    4. Beckmann CF
    5. Behrens TEJ
    6. Johansen-Berg H
    7. Bannister PR
    8. De Luca M
    9. Drobnjak I
    10. Flitney DE
    11. Niazy RK
    12. Saunders J
    13. Vickers J
    14. Zhang Y
    15. De Stefano N
    16. Brady JM
    17. Matthews PM

    (2004) Advances in functional and structural mr image analysis and implementation as fsl

    NeuroImage 23 Suppl 1:S208–S219.

    https://doi.org/10.1016/j.neuroimage.2004.07.051

    • PubMed
    • Google Scholar
55. 1. Soutschek A
    2. Ruff CC
    3. Strombach T
    4. Kalenscher T
    5. Tobler PN

    (2016) Brain stimulation reveals crucial role of overcoming self-centeredness in self-control

    Science Advances 2:e1600992.

    https://doi.org/10.1126/sciadv.1600992

    • PubMed
    • Google Scholar
56. 1. Suzuki S
    2. Harasawa N
    3. Ueno K
    4. Gardner JL
    5. Ichinohe N
    6. Haruno M
    7. Cheng K
    8. Nakahara H

    (2012) Learning to simulate others’ decisions

    Neuron 74:1125–1137.

    https://doi.org/10.1016/j.neuron.2012.04.030

    • PubMed
    • Google Scholar
57. 1. Trudel N
    2. Scholl J
    3. Klein-Flügge MC
    4. Fouragnan E
    5. Tankelevitch L
    6. Wittmann MK
    7. Rushworth MFS

    (2021) Polarity of uncertainty representation during exploration and exploitation in ventromedial prefrontal cortex

    Nature Human Behaviour 5:83–98.

    https://doi.org/10.1038/s41562-020-0929-3

    • PubMed
    • Google Scholar
58. 1. Will GJ
    2. Rutledge RB
    3. Moutoussis M
    4. Dolan RJ

    (2017) Neural and computational processes underlying dynamic changes in self-esteem

    eLife 6:e28098.

    https://doi.org/10.7554/eLife.28098

    • PubMed
    • Google Scholar
59. 1. Wittmann MK
    2. Kolling N
    3. Faber NS
    4. Scholl J
    5. Nelissen N
    6. Rushworth MFS

    (2016) Self-other mergence in the frontal cortex during cooperation and competition

    Neuron 91:482–493.

    https://doi.org/10.1016/j.neuron.2016.06.022

    • PubMed
    • Google Scholar
60. 1. Wittmann MK
    2. Lockwood PL
    3. Rushworth MFS

    (2018) Neural mechanisms of social cognition in primates

    Annual Review of Neuroscience 41:99–118.

    https://doi.org/10.1146/annurev-neuro-080317-061450

    • PubMed
    • Google Scholar
61. 1. Wittmann MK
    2. Fouragnan E
    3. Folloni D
    4. Klein-Flügge MC
    5. Chau BKH
    6. Khamassi M
    7. Rushworth MFS

    (2020) Global reward state affects learning and activity in raphe nucleus and anterior insula in monkeys

    Nature Communications 11:3771.

    https://doi.org/10.1038/s41467-020-17343-w

    • PubMed
    • Google Scholar
62. 1. Wittmann MK
    2. Trudel N
    3. Trier HA
    4. Klein-Flügge MC
    5. Sel A
    6. Verhagen L
    7. Rushworth MFS

    (2021) Causal manipulation of self-other mergence in the dorsomedial prefrontal cortex

    Neuron 109:2353–2361.

    https://doi.org/10.1016/j.neuron.2021.05.027

    • PubMed
    • Google Scholar
63. 1. Yoshida K
    2. Saito N
    3. Iriki A
    4. Isoda M

    (2012) Social error monitoring in macaque frontal cortex

    Nature Neuroscience 15:1307–1312.

    https://doi.org/10.1038/nn.3180

    • PubMed
    • Google Scholar
64. Conference

    1. Yu AJ
    2. Cohen JD

    (2008)

    Sequential effects: superstition or rational behavior?

    Advances in Neural Information Processing Systems. pp. 1873–1880.

    • Google Scholar
65. 1. Zink CF
    2. Tong Y
    3. Chen Q
    4. Bassett DS
    5. Stein JL
    6. Meyer-Lindenberg A

    (2008) Know your place: neural processing of social hierarchy in humans

    Neuron 58:273–283.

    https://doi.org/10.1016/j.neuron.2008.01.025

    • PubMed
    • Google Scholar

Decision letter after peer review:

Thank you for submitting your article "Neural activity tracking identity and confidence in social information" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Michael Frank as the Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1) It is not just the "framing" of information as social that differs across conditions, but also the stimuli (i.e., faces vs. fruit). Ideally, the behavioral experiment should be performed again, with the same cover-story, but using abstract shapes or symbols as stimuli, to rule out the possibility that the complexity/familiarity/salience of faces elicit attentional processes that account for the results. If additional data collection is not possible, the authors need to make a compelling case for why the results can be interpreted despite this confound.

2) Make sure to place the main question (i.e., whether humans learn differently from social vs. non-social information) in the context of relevant previous neural and computational work, both in the Introduction and Discussion, and identify what specific novel conceptual insights can be gleaned from these particular results (see the reviewers' comments below for relevant papers).

3) Use model simulations to test some of the explanations proposed in the Discussion (e.g., does changing the learning index of the Bayesian model produce the observed confidence intervals?).

4) Please perform an MVPA searchlight to assess the specificity of the effects in dmPFC and TPJ.

5) Behavioural indices and analyses need to be better justified (e.g., why should the interval size exactly match the angular error?).

6) Please provide a full description of the computational model in the main text.

7) Carefully consider and respond to all the reviewers' comments, appended below. In your reply, list each specific reviewer comment, followed by the response and the relevant revision/rebuttal.

Reviewer #1 (Recommendations for the authors):

The authors should provide the full description of the computational model (rather than the schematic illustration).

For the GLM in Figure 2e, I would suggest the authors use the mixed-effect model.

Reviewer #2 (Recommendations for the authors):

Regarding point 1 in the public review: I would be happy if the authors simply engaged more deeply with prior theoretical accounts, both in the introduction and in the discussion. However, I do think the theoretical contributions of this paper could be strengthened by building on some of the proposals in the discussion. In the discussion, the authors propose a few reasons why participants' estimates may have been less influenced by subjective uncertainty in the social condition. In particular, participants could have stronger prior expectations about the stability of social sources, or they could show a steeper decline in uncertainty over time. (Though this latter proposal seems contradicted by the results – wouldn't you then expect to see a larger "learning index" after the first update?) The Bayesian model used in this paper could be used to demonstrate the plausibility of these proposals – if you simulate the model forward using different priors or learning rates, does the model capture qualitative patterns in human behavior?

Regarding point 2: It would be informative to pair the multivariate analysis with a whole-brain searchlight, to test whether any other regions show this effect.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Neural activity tracking identity and confidence in social information" for further consideration by eLife. Your revised article has been evaluated by Michael Frank (Senior Editor), Mimi Liljeholm (Reviewing Editor), and two expert reviewers.

The manuscript has been greatly improved and the only remaining request is that you characterize the additional analyses, prompted by the first round of reviews, as exploratory, since they were not planned in the original submission.

Reviewer #1 (Recommendations for the authors):

The authors have adequately addressed all the concerns.

Reviewer #2 (Recommendations for the authors):

I thank the authors for their thoughtful and thorough responses to reviewer comments. My past review raised two key concerns. First, I suggested that the paper could engage more deeply with past work on the computational basis of social learning, in order to ground their discussion on the differences between social and non-social learning. Second, I raised the concern that condition differences between social and non-social stimuli could be driven by lower-level features, such as the attentional salience or visual distinctiveness of the stimuli. Overall, I believe that the authors have gone above and beyond to address these concerns.

First, the authors revised their introduction and discussion to cite past scholarship in social learning. But, beyond that, they also substantially expanded their analyses to disentangle differences between the underlying mechanisms driving behavior in social and non-social conditions. The authors found that participants did not have different prior expectations about the performance of social vs. non-social predictors. Instead, using a modified, noisy Bayesian model, the authors suggest that differences between conditions may be driven by degrees of uncertainty in the belief update and that greater noise in the update process leads to larger changes in confidence across different encounters with the predictor. I agree that these analyses have significantly improved the manuscript and made a new theoretical contribution in their own right, as they suggest a common computational mechanism underlying observed behavioral differences between conditions.

If I can offer a small suggestion: Because these analyses and hypotheses were not planned in the initial submission, it would be helpful to specify that they are exploratory/an extension when they are first presented in the main text (p. 10-11).

Second, the authors added several control analyses as a supplement. In order to test whether condition differences could be accounted for by attentional engagement, the authors inspected the constant of the GLM. They found no condition differences unaccounted for by the parametric regressors included in the GLM, either in attentional regions or in the regions that were the focus of their study. Next, in order to test whether social and non-social stimuli differed in their visual distinctiveness, the authors used RSA to compare the EDI between conditions in early visual areas. Again, the authors found no difference, suggesting that observed condition differences are unlikely to be driven by these lower-level features.

Reviewer #1 (Recommendations for the authors):

The authors should provide the full description of the computational model (rather than the schematic illustration).

We have now integrated a full description of the Bayesian model into the Methods section.

Bayesian model. We used a Bayesian model ¹⁸ (Figure 2 —figure supplement 2) to guide our analyses. We focussed, first, on how participants’ behaviour might differ between the social and non-social situations and, second, on how participants estimated their uncertainty about the predictions which was revealed by their setting of the confidence interval. We used a Bayesian model to estimate participants’ beliefs about the performance of a social and non-social advisor/predictor. The model was fitted to each social and non-social advisor/predictor to derive belief estimates about the distribution that determined the observed angular error, i.e. the distance between the social and non-social target estimate and the true target location. Angular errors were drawn from a distribution that varied according to the value of sigma (σ) which refers to the standard deviation of the normal distribution. Across trials, participants observe how well each predictor performs and thereby participants form a belief about the average angular error, i.e. the estimated σ (referring to sigma-hat (${\displaystyle \hat{\sigma }}$), reflecting an estimate rather than the true value of sigma) of the distribution.

According to Bayes’ rule, a belief is updated by multiplication of a prior belief and a likelihood distribution resulting in a posterior belief, i.e. the belief update (Figure 2d). We assumed a uniform prior across parameter space of σ-hat. In other words, every possible σ value could underlie the predictor’s distribution with equal probability.

${\displaystyle p\left(\hat{\sigma }\right)=U(1,140)}$

(2)

The likelihood function assumed a normal distribution of the observed angular error (x), and described the probability of the observation x given each possible sigma value:

${\displaystyle p\left(x|\hat{\sigma }\right)=N\left(x|\mu =0,\hat{\sigma }\right).}$

(3)

With Bayes rule, a trial wise posterior distribution proportional to the multiplication of the likelihood and prior distribution:

${\displaystyle p\left(\hat{\sigma }|x\right)p\left(x|\hat{\sigma }\right)p(\hat{\sigma })}$

(4)

where p (${\displaystyle \hat{\sigma }}$) is the prior distribution, p(x | ${\displaystyle \hat{\sigma }}$) is the likelihood function and p(${\displaystyle \hat{\sigma }}$| x), is the posterior pdf across parameter space. The posterior pdf is the updated belief across σ space and is used as prior for the next trial of the same predictor.

Each posterior was normalised to ensure that probabilities across all σ values added up to one:

${\displaystyle p\left(\hat{\sigma }|x\right)=\frac{p\hat{\sigma }(x)}{\mathrm{\Sigma }p(\hat{\sigma }x)}}$

(5)

We submitted the computational variables from our model to a general linear model (GLM) analysis to predict participants’ interval setting. For every trial, we extracted two model based subjective belief parameters that captured two key features of each participant’s likely beliefs about a predictor’s performance. First, the belief of how accurately a predictor would predict the target was captured by the mode of the probability density function on each trial associated with the selected predictor:

accuracy = max [p (${\displaystyle \hat{\sigma }}$)] * (-1)

(6)

Note that a higher value of max [p (${\displaystyle \hat{\sigma }}$)] corresponds to a larger angular error; in other words, it corresponds to a larger deviation between the true target location and the predicted target location. To derive an intuitive interpretation, the sign of max [p (${\displaystyle \hat{\sigma }}$)] was reversed such that now a greater positive value can be interpreted as higher accuracy. In sum, the accuracy estimates as defined in equation (5) represents a point-estimate of a belief distribution in σ-hat (${\displaystyle \hat{\sigma }}$).

A second model parameter was derived to reflect a participant’s subjective uncertainty in that accuracy belief. On each trial, a percentage (2.5%) of the lower and upper tail of the prior distribution around the estimated σ value was derived. The difference between σ values at each of the tails was subtracted to derive the estimated ‘uncertainty’ variable:

${\displaystyle \hat{\sigma }}$_{high ⟵} cumulative (p (${\displaystyle \hat{\sigma }}$)) = 97.5%

${\displaystyle \hat{\sigma }}$_{low ⟵} cumulative (p (${\displaystyle \hat{\sigma }}$)) = 2.5% uncertainty = ${\displaystyle \hat{\sigma }}$_high – ${\displaystyle \hat{\sigma }}$_low

(7)

Please refer to the Figure 2 —figure supplement 2 for a schematic illustration of the Bayesian model.

For the GLM in Figure 2e, I would suggest the authors use the mixed-effect model.

Thank you for this suggestion. We agree that a mixed-effect model is, indeed, a sensitive approach to compare the differences between social and non-social conditions as they take the data dependence into account¹⁹. In the behavioural GLM (Figure 2e), we showed that there are differences in the impact of Bayesian uncertainty estimates on trial-wise confidence judgments about a predictor’s performance: participants were less impacted by the uncertainty around the accuracy estimate when making confidence judgments about social advisors compared to non-social cues. Hence, the aim of the mixed-effect model is mainly to test whether such a difference in the ‘predictor uncertainty’ effect can be replicated. We used a random slope mixed-effect model (lme function in matlab).

The random slope mixed-effect model of interest includes all regressors of interest as fixed effects (Figure 2e and Figure 1 —figure supplement 2), random slopes for each fixed effect as well as a random intercept. We additionally included a binary variable (condition_ID) to denote the social (condition=1) and non-social (condition=0) condition. All variables were normalised. The effect of interest is the interaction effect between the condition variable and the fixed effect ‘predictor uncertainty’.

To evaluate the random slope mixed-effect model, we compared two models – one model that included all regressors of interest [H-alternative] to a model that included all regressors except the effect of interest ‘predictor uncertainty’ effect and the interaction ‘predictor uncertainty x groupID’ [H-null]. Then, we can compare the model fit and evaluate whether the addition of uncertainty variables is meaningful. Noteworthy, when doing such a random slope mixed-effect model comparison, it is crucial to keep the ‘predictor uncertainty’ variable as random slope in both models. In the syntax of the lme function, the models were specified as follows:

H-null:

Confidence ~ (predictor accuracy + payoff_t-1 + confidence_t-1)* conditionID + ((predictor accuracy + predictor uncertainty + payoff_t-1 + confidence_t-1)*conditionID | subjects).

H-alternative:

Confidence ~ (predictor accuracy + predictor uncertainty + payoff_t-1 + confidence_t-1)* conditionID + ((predictor accuracy + predictor uncertainty + payoff_t-1 + confidence_t1)*conditionID | subjects).

First, a likelihood ratio test was used to compare the goodness of fit of the two competing models. The alternative model that included the additional ‘predictor uncertainty’ variable as fixed effect and its interaction with the conditionID was a better model of the data (χ2(2) = 22.238, p=.483e-05). Next, we tested whether there was a significant interaction of ‘uncertainty predictor x conditionID. When using a random slope mixed-effect model, we replicated the difference between conditions (‘uncertainty predictor x conditionID, estimate = -0.178, SE = 0.07, p = 0.017) – again showing a difference in the impact of subjective uncertainty when making confidence judgments about social advisors or non-social cues.

We included the results in the main text and Supplementary material as follows:

Main text, page 9:

“However, when judging non-social predictors, but not social predictors, interval settings were additionally widened as a function of subjective uncertainty: participants set a larger interval when they were more uncertain about the non-social cue’s accuracy to predict the target (paired t-test: social vs. non-social, t(23) = 3.5, p=0.0018, 95% CI = [0.11 0.42]; one sample t-test: non-social, t(23) = -3.9, p<0.0001, 95% CI = [-0.35 -0.1]; one sample t-test: social, t(23)=0.6, p=0.56, 95% CI = [-0.1 0.19], Figure 2e). The difference between social and non-social conditions was replicated when using a random slope mixed effect model (Figure 2 —figure supplement 1, legend). Uncertainty about the accuracy of social advisors, however, did not have the same impact: participants’ interval settings for the social advisors appeared only to be guided by the peak of the Bayesian probability distribution, as indicated by the accuracy effect.”

Reviewer #2 (Recommendations for the authors):

Regarding point 1 in the public review: I would be happy if the authors simply engaged more deeply with prior theoretical accounts, both in the introduction and in the discussion. However, I do think the theoretical contributions of this paper could be strengthened by building on some of the proposals in the discussion. In the discussion, the authors propose a few reasons why participants' estimates may have been less influenced by subjective uncertainty in the social condition. In particular, participants could have stronger prior expectations about the stability of social sources, or they could show a steeper decline in uncertainty over time. (Though this latter proposal seems contradicted by the results – wouldn't you then expect to see a larger "learning index" after the first update?) The Bayesian model used in this paper could be used to demonstrate the plausibility of these proposals – if you simulate the model forward using different priors or learning rates, does the model capture qualitative patterns in human behavior?

We thank the reviewer for this suggestion. In response, apart from engaging more deeply with prior theoretical accounts, we have undertaken further analyses that aim to disentangle the underlying mechanisms that might differ between both social and nonsocial conditions might provide additional theoretical contributions. We followed the advice of the review and show additional model simulations and analyses that aim to disentangle the differences in more detail. These new results allowed clearer interpretations to be made.

In the current study, we show that judgments made about non-social predictors were changed more strongly as a function of the subjective uncertainty: participants set a larger interval, indicating lower confidence, when they were more uncertain about the non-social cue’s accuracy to predict the target. In response to the reviewer’s comments, the new analyses were aimed at understanding under which conditions such a negative uncertainty effect might emerge.

Prior expectations of performance

Related to the first point of the reviewer, we compared whether participants had different prior expectations in the social condition compared to the non-social condition. One way to compare prior expectations is by comparing the first interval set for each advisor/predictor. In such a way, we test whether the prior beliefs before observing any social or non-social information differ between conditions. Even though this does not test the impact of prior expectations on subsequent belief updates, it does test whether participants have generally different expectations about the performance of social advisors or non-social predictors. There was no difference in this measure between social or non-social cues (Author response image 1; paired t-test social vs. non-social, t(23) = 0.01, p=0.98, 95% CI = [-0.067 0.68]).

Author response image 1

Download asset Open asset

Learning across time

We have now seen that participants do not have an initial bias when predicting performances in social or non-social conditions. This suggests that differences between conditions might emerge across time when encountering predictors multiple times. We tested whether inherent differences in how beliefs are updated according to new observations might result in different impacts of uncertainty on interval setting between social and non-social conditions. More specifically, we tested whether the integration of new evidence differed between social and non-social conditions; for example, recent observations might be weighted more strongly for non-social cues while past observations might be weighted more strongly for social cues. This approach was inspired by the reviewer’s comments about potential differences in the speed of learning as well as the reduction of uncertainty with increasing predictor encounters. Similar ideas were tested in previous studies, when comparing the learning rate (i.e. the speed of learning) in environments of different volatilities^12,13. In these studies, a smaller learning rate was prevalent in stable environments during which reward rates change slower over time, while higher learning rates often reflect learning in volatile environments so that recent observations have a stronger impact on behaviour. Even though most studies derived these learning rates with reinforcement learning models, similar ideas can be translated into a Bayesian model. For example, an established way of changing the speed of learning in a Bayesian model is to introduce noise during the update process¹⁴. This noise is equivalent to adding in some of the initial prior distribution and this will make the Bayesian updates more flexible to adapt to changing environments. It will widen the belief distribution and thereby make it more uncertain. Recent information has more weight on the belief update within a Bayesian model when beliefs are uncertain. This increases the speed of learning. In other words, a wide distribution (after adding noise) allows for quick integration of new information. On the contrary, a narrow distribution does not integrate new observations as strongly and instead relies more heavily on previous information; this corresponds to a small learning rate. So, we would expect a steep decline of uncertainty to be related to a smaller learning index while a slower decline of uncertainty is related to a larger learning index. We hypothesized that participants reduce their uncertainty quicker when observing social information, thereby anchoring more strongly on previous beliefs instead of integrating new observations flexibly. Vice versa, we hypothesized a less steep decline of uncertainty when observing non-social information, indicating that new information can be flexibly integrated during the belief update (new Figure 3a).

We modified the original Bayesian model (Figure 2d, Figure 2 —figure supplement 2) by adding a uniform distribution (equivalent to our prior distribution) to each belief update – we refer to this as noise addition to the Bayesian model^14,21. We varied the amount of noise between δ = [0,1], while δ = 0 equals the original Bayesian model and δ = 1 represents a very noisy Bayesian model. The uniform distribution was selected to match the first prior belief before any observation was made (equation 2). This δ range resulted in a continuous increase of subjective uncertainty around the belief about the angular error (Figure 3b-c). The modified posterior distribution denoted as p′(${\displaystyle \hat{\sigma }}$| x) was derived at each trial as follows:

p′($\hat{\sigma }$| x) = p (${\displaystyle \hat{\sigma }}$| x) + (U(1,140) ∗ δ)

(10)

The noisy posterior distribution as ${\displaystyle p^{\mathrm{\prime }}(\hat{\sigma }|x)}$ was normalised as follows:

${\displaystyle p^{\mathrm{\prime }}(\hat{\sigma }|x)=\frac{p^{\mathrm{\prime }}(\hat{\sigma }|x)}{\sum p^{\mathrm{\prime }}(\hat{\sigma }|x)}}$

(11)

We applied each noisy Bayesian model to participants’ choices within the social and nonsocial condition.

The addition of a uniform distribution changed two key features of the belief distribution: first, the width of the distribution remains larger with additional observations, thereby making it possible to integrate new observations more flexibly. To show this more clearly, we extracted the model-derived uncertainty estimate across multiple encounters of the same predictor for the original model and the fully noisy Bayesian model (Figure 3 —figure supplement 1a). The model-derived ‘uncertainty estimate’ of a noisy Bayesian model decays more slowly compared to the ‘uncertainty estimate’ of the original Bayesian model (upper panel). Second, the model-derived ‘accuracy estimate’ reflects more recent observations in a noisy Bayesian model compared to the ‘accuracy estimate’ derived from the original Bayesian model, which integrates past observations more strongly (lower panel). Hence, as mentioned beforehand, a rapid decay of uncertainty implies a small learning index; or in other words, stronger integration of past compared to recent observations.

In the following analyses, we tested whether an increasingly noisy Bayesian model mimics behaviour that is observed in the non-social compared to social condition. For example, we tested whether an increasingly noisy Bayesian model also exhibits a strongly negative ‘predictor uncertainty’ effect on interval setting (Figure 2e). In such a way, we can test whether differences in noise in the updating process of a Bayesian model might reproduce differences in learning-related behaviour seen in the social and non-social conditions.

We used these modified Bayesian models to simulate trial-wise interval setting for each participant according to the observations they made when selecting a particular advisor or non-social cue. We simulated interval setting at each trial and examined whether an increase in noise produced model behaviours that resembled participant behaviour patterns observed in the non-social condition as opposed to social condition. At each trial, we used the accuracy estimate (Methods, equation 6) – which represents a subjective belief about a single angular error -- to derive an interval setting for the selected predictor. To do so, we first derived the point-estimate of the belief distribution at each trial (Methods, equation 6) and multiplied it with the size of one interval step on the circle. The step size was derived by dividing the circle size by the maximum number of possible steps. Here is an example of transforming an accuracy estimate into an interval: let’s assume the belief about the angular error at the current trial is 50 (Methods, equation 6). Now, we are trying to transform this number into an interval for the current predictor on a given trial. To obtain the size of one interval step, the circle size (360 degrees) is divided by the maximum number of interval steps (40 steps; note, 20 steps on each side), which results in nine degrees that represents the size of one interval step. Next, the accuracy estimate in radians (0,87) is multiplied by the step size in radians (0,1571) resulting in an interval of 0,137 radians or 7,85 degrees. The final interval size would be 7,85.

We repeated behavioural analyses (Figure 2b,e,f) to test whether intervals derived from more noisy Bayesian models mimic intervals set by participants in the non-social condition: greater changes in interval setting across trials (Figure 3 —figure supplement 1b), a negative ‘predictor uncertainty' effect on interval setting (Figure 3 —figure supplement 1c), and a higher learning index (Figure 3d).

First, we repeated the most crucial analysis -- the linear regression analysis (Figure 2e) and hypothesized that intervals that were simulated from noisy Bayesian models would also show a greater negative ‘predictor uncertainty’ effect on interval setting. This was indeed the case: irrespective of social or non-social conditions, the addition of noise (increased weighting of the uniform distribution in each belief update) led to an increasingly negative ‘predictor uncertainty’ effect on confidence judgment (new Figure 3d). In Figure 3d, we show the regression weights (y-axis) for the ‘predictor uncertainty’ on confidence judgment with increasing noise (x-axis).

This new finding extends the results and suggests a theoretical account of the behavioural differences between social and non-social conditions. Increasing the noise of the belief update mimics behaviour that is observed in the non-social condition: an increasingly negative effect of ‘predictor uncertainty’ on confidence judgment. Noteworthily, there was no difference in the impact that the noise had in the social and non-social conditions, thereby suggesting that an increase in the Bayesian noise leads to an increasingly negative impact of ‘predictor uncertainty’ on confidence judgments irrespective of the condition. Hence, we can conclude that different degrees of uncertainty within the belief update is a reasonable explanation that can underlie the differences observed between social and nonsocial conditions.

Next, we used these simulated confidence intervals and repeated the descriptive behavioural analyses to test whether interval settings that were derived from more noisy Bayesian models mimic behavioural patterns observed in non-social compared to social conditions. For example, more noise in the belief update should lead to more flexible integration of new information and hence should potentially lead to a greater change of confidence judgments across predictor encounters (Figure 2b). Further, a greater reliance on recent information should lead to prediction errors more strongly in the next confidence judgment; hence, it should result in a higher learning index (Figure 2f). We used the simulated confidence interval from Bayesian models on a continuum of noise integration (i.e. different weighting of the uniform distribution into the belief update) and derived again both absolute confidence change and learning indices (Figure 3 —figure supplement 1b-c).

‘Absolute confidence change’ and ‘learning index’ increase with increasing noise weight, thereby mimicking the difference between social and non-social conditions. Further, these analyses demonstrate the tight relationship between descriptive analyses and model-based analyses thereby reinforcing both type of analyses.

We thank the reviewer for making this point, as we believe that these additional analyses allow theoretical inferences to be made in a more direct manner; we think that it has significantly contributed towards a deeper understanding of the mechanisms involved in the social and non-social conditions. Further, it provides a novel account of how we make judgments when being presented with social and non-social information.

We made substantial changes to the main text, figures and supplementary material to include these changes:

Main text, page 10-11 new section:

“The impact of noise in belief updating in social and non-social conditions

So far, we have shown that, in comparison to non-social predictors, participants changed their interval settings about social advisors less drastically across time, relied on observations made further in the past, and were less impacted by their subjective uncertainty when they did so (Figure 2). […] Correspondingly, in the non-social condition, participants express more uncertainty in their confidence estimate when they set the interval, and they also change their beliefs more quickly. Therefore, noisy Bayesian updating can account for key differences between social and non-social condition.”

Methods, page 23 new section:

“Extension of Bayesian model with varying amounts of noise

We modified the original Bayesian model (Figure 2d, Figure 2 —figure supplement 2) to test whether the integration of new evidence differed between social and non-social conditions; for example, recent observations might be weighted more strongly for non-social cues while past observations might be weighted more strongly for social cues. We tested these ideas using established ways of changing the speed of learning during Bayesian updates^14,21. For example, an established way of changing the speed of learning in a Bayesian model is to introduce noise during the update process¹⁴. This noise is equivalent to adding in some of the initial prior distribution and this will make the Bayesian updates more flexible to adapt to changing environments. We added a uniform distribution to each belief update; we refer to this as noise addition to the Bayesian model. We varied the amount of noise between δ = [0,1], while δ = 0 equals the original Bayesian model and δ = 1 represents a very noisy Bayesian model. The uniform distribution was selected to match the first prior belief before any observation was made (equation 2). This δ range resulted in a continuous increase of subjective uncertainty around the belief about the angular error (Figure 3b-c).

The modified posterior distribution denoted as p′(${\displaystyle \hat{\sigma }}$| x) was derived at each trial as follows:

p′(${\displaystyle \hat{\sigma }}$| x) = p(${\displaystyle \hat{\sigma }}$| x) + (U(1,140) ∗ δ)

(10)

The noisy posterior distribution as p′(${\displaystyle \hat{\sigma }}$| x) was normalised as follows:

${\displaystyle p^{\mathrm{\prime }}(\hat{\sigma }|x)=\frac{p^{\mathrm{\prime }}(\hat{\sigma }|x)}{\sum p^{\mathrm{\prime }}(\hat{\sigma }|x)}}$

We applied each noisy Bayesian model to participants’ interval settings in the social and non-social conditions. Then, we used the model to simulate interval setting at each trial with varying noise levels and then examined whether an increase in noise results in simulation behaviour that resembles behavioural patterns observed in the non-social conditions that were different to behavioural patterns observed in the social condition. At each trial, we used the accuracy estimate (equation 6) to derive a confidence interval associated with the selected predictor. We transformed the accuracy estimate – which represents a subjective belief about a single angular error – into an interval setting by multiplying the trial specific angular error with the step size. The step size was derived by dividing the circle size by the maximum number of possible steps. Here is an example of transforming an accuracy estimate into an interval: let’s assume the belief about the angular error at the current trial is 50 (Methods, equation 6). Now, we are trying to transform this number into an interval for the current predictor on a given trial. To obtain the size of one interval step, the circle size (360 degrees) is divided by the maximum number of interval steps (40 steps; note, 20 steps on each side), which results in nine degrees that represents the size of one interval step. Next, the accuracy estimate in radians (0,87) is multiplied by the step size in radians (0,1571) resulting in an interval of 0,137 radians or 7,85 degrees. The final interval size would be 7,85.”

We repeated behavioural analyses (Figure 2b,e,f) to test whether confidence intervals derived from more noisy Bayesian models mimic behavioural patterns observed in the nonsocial condition: greater changes of confidence across trials (Figure 3 —figure supplement 1b), a greater negative ‘predictor uncertainty' on confidence judgment (Figure 3 —figure supplement 1c) and a greater learning index (Figure 3d).”

Discussion, page 14:

“It may be because we make just such assumptions that past observations are used to predict performance levels that people are likely to exhibit next ^15,16. An alternative explanation might be that participants experience a steeper decline of subjective uncertainty in their beliefs about the accuracy of social advice, resulting in a narrower prior distribution, during the next encounter with the same advisor. We used a series of simulations to investigate how uncertainty about beliefs changed from trial to trial and showed that belief updates about non-social cues were consistent with a noisier update process that diminished the impact of experiences over the longer term. From a Bayesian perspective, greater certainty about the value of advice means that contradictory evidence will need to be stronger to alter one’s beliefs. In the absence of such evidence, a Bayesian agent is more likely to repeat previous judgments. Just as in a confirmation bias ¹⁷, such a perspective suggests that once we are more certain about others’ features, for example, their character traits, we are less likely to change our opinions about them.”

Regarding point 2: It would be informative to pair the multivariate analysis with a whole-brain searchlight, to test whether any other regions show this effect.

We have complied with the request and conducted a whole-brain searchlight analysis. To implement the searchlight analysis, we used SPM12, RSA toolbox²² and custom based scripts while closely following analysis steps of previously reported whole-brain searchlight analyses²³. We conducted the whole-brain analysis in 16 participants, because of technical difficulties matching the data from FSL to SPM. In accordance with previous studies²³ and best-practices the searchlight analysis was based on smoothed data and multivariate noise normalization was applied to the voxel activity pattern to improve reliability²². The searchlight sphere was 15mm (approximately 100 voxels).

We calculated the Euclidian distance between all conditions. Next, for each condition, we calculated the Exemplar Discriminability Index (EDI) by subtracting the average Euclidian distances between neural activation maps of the same predictors from the average distances between bundles of different predictors (while taking care not to include the distance between two instances of precisely the same bundle). The resulting whole-brain maps were smoothed at 4mm.

Comparison analyses between social and non-social conditions were run on a second-level ([-1 1] for social > non-social and [1 -1] for social < non-social) to reveal the areas that responded specifically to one condition. Note the sign of each of the contrasts: we were interested in negative activation patterns, as a negative EDI is expected when

representations of the same cues are less dissimilar than representations of different cues (Figure 4g for a schematic explanation of EDI). However, we did not find a whole-brain cluster corrected result. We report this in a new supplementary figure and show subthreshold activation maps. Strong signals are apparent near our a priori ROI. Note that a comparison between the previous ROI-based analysis and the whole-brain searchlight analysis cannot be directly made and that the latter should rather be treated as a complementary analysis²³. This is for the following reasons. The searchlight sphere comprises 100 voxels. Its spherical shape does not respect precise anatomical boundaries such as for example our pTPJ mask. Note that the dmPFC and pTPJ masks contained on average nine times the number of voxels relative to the searchlight sphere. Therefore, as in previous work (e.g., Bang et al., 2020, Nat Commun), our whole brain searchlight results are well in line with our ROI results, but they differ in degree, because of the size and shape of the neural masks used to assess neural similarity.

In addition to including this new Figure 4 —figure supplement 3, we have made corresponding changes to the main text.

Main text, section: dmPFC and pTPJ covary with social confidence and encode social advisor identity

“An exploratory whole-brain searchlight analysis was performed to test the specificity of dmPFC and pTPJ encoding the identity of social compared to non-social cues. No other cluster-corrected effects were found (Figure 4 —figure supplement 3). Additional control analyses show that neural differences between social and non-social conditions are not due to the visually different set of stimuli used in the experiment but are instead representing fundamental differences in processing social compared to non-social information (Figure 4 —figure supplement 4).”

References

1. Heyes, C. (2012). What’s social about social learning? Journal of Comparative Psychology 126, 193–202. 10.1037/a0025180.

2. Chang, S.W.C., and Dal Monte, O. (2018). Shining Light on Social Learning Circuits. Trends in Cognitive Sciences 22, 673–675. 10.1016/j.tics.2018.05.002.

3. Diaconescu, A.O., Mathys, C., Weber, L.A.E., Kasper, L., Mauer, J., and Stephan, K.E. (2017). Hierarchical prediction errors in midbrain and septum during social learning. Soc Cogn Affect Neurosci 12, 618–634. 10.1093/scan/nsw171.

4. Frith, C., and Frith, U. (2010). Learning from Others: Introduction to the Special Review Series on Social Neuroscience. Neuron 65, 739–743. 10.1016/j.neuron.2010.03.015.

5. Frith, C.D., and Frith, U. (2012). Mechanisms of Social Cognition. Annu. Rev. Psychol. 63, 287–313. 10.1146/annurev-psych-120710-100449.

6. Grabenhorst, F., and Schultz, W. (2021). Functions of primate amygdala neurons in economic decisions and social decision simulation. Behavioural Brain Research 409, 113318. 10.1016/j.bbr.2021.113318.

7. Lockwood, P.L., Apps, M.A.J., and Chang, S.W.C. (2020). Is There a ‘Social’ Brain? Implementations and Algorithms. Trends in Cognitive Sciences, S1364661320301686. 10.1016/j.tics.2020.06.011.

8. Soutschek, A., Ruff, C.C., Strombach, T., Kalenscher, T., and Tobler, P.N. (2016). Brain stimulation reveals crucial role of overcoming self-centeredness in self-control. Sci. Adv. 2, e1600992. 10.1126/sciadv.1600992.

9. Wittmann, M.K., Lockwood, P.L., and Rushworth, M.F.S. (2018). Neural Mechanisms of Social Cognition in Primates. Annu. Rev. Neurosci. 41, 99–118. 10.1146/annurev-neuro080317-061450.

10. Shafto, P., Goodman, N.D., and Frank, M.C. (2012). Learning From Others: The Consequences of Psychological Reasoning for Human Learning. Perspect Psychol Sci 7, 341– 351. 10.1177/1745691612448481.

11. McGuire, J.T., Nassar, M.R., Gold, J.I., and Kable, J.W. (2014). Functionally Dissociable Influences on Learning Rate in a Dynamic Environment. Neuron 84, 870–881. 10.1016/j.neuron.2014.10.013.

12. Behrens, T.E.J., Woolrich, M.W., Walton, M.E., and Rushworth, M.F.S. (2007). Learning the value of information in an uncertain world. Nature Neuroscience 10, 1214– 1221. 10.1038/nn1954.

13. Meder, D., Kolling, N., Verhagen, L., Wittmann, M.K., Scholl, J., Madsen, K.H., Hulme, O.J., Behrens, T.E.J., and Rushworth, M.F.S. (2017). Simultaneous representation of a spectrum of dynamically changing value estimates during decision making. Nat Commun 8, 1942. 10.1038/s41467-017-02169-w.

14. Allenmark, F., Müller, H.J., and Shi, Z. (2018). Inter-trial effects in visual pop-out search: Factorial comparison of Bayesian updating models. PLoS Comput Biol 14, e1006328. 10.1371/journal.pcbi.1006328.

15. Wittmann, M., Trudel, N., Trier, H.A., Klein-Flügge, M., Sel, A., Verhagen, L., and Rushworth, M.F.S. (2021). Causal manipulation of self-other mergence in the dorsomedial prefrontal cortex. Neuron.

16. Wittmann, M.K., Kolling, N., Faber, N.S., Scholl, J., Nelissen, N., and Rushworth, M.F.S. (2016). Self-Other Mergence in the Frontal Cortex during Cooperation and Competition. Neuron 91, 482–493. 10.1016/j.neuron.2016.06.022.

17. Kappes, A., Harvey, A.H., Lohrenz, T., Montague, P.R., and Sharot, T. (2020). Confirmation bias in the utilization of others’ opinion strength. Nat Neurosci 23, 130–137. 10.1038/s41593-019-0549-2.

18. Trudel, N., Scholl, J., Klein-Flügge, M.C., Fouragnan, E., Tankelevitch, L., Wittmann, M.K., and Rushworth, M.F.S. (2021). Polarity of uncertainty representation during exploration and exploitation in ventromedial prefrontal cortex. Nat Hum Behav. 10.1038/s41562-020-0929-3.

19. Yu, Z., Guindani, M., Grieco, S.F., Chen, L., Holmes, T.C., and Xu, X. (2022). Beyond t test and ANOVA: applications of mixed-effects models for more rigorous statistical analysis in neuroscience research. Neuron 110, 21–35. 10.1016/j.neuron.2021.10.030.

20. Mars, R.B., Jbabdi, S., Sallet, J., O’Reilly, J.X., Croxson, P.L., Olivier, E., Noonan, M.P., Bergmann, C., Mitchell, A.S., Baxter, M.G., et al. (2011). Diffusion-Weighted Imaging Tractography-Based Parcellation of the Human Parietal Cortex and Comparison with Human and Macaque Resting-State Functional Connectivity. Journal of Neuroscience 31, 4087– 4100. 10.1523/JNEUROSCI.5102-10.2011.

21. Yu, A.J., and Cohen, J.D. Sequential effects: Superstition or rational behavior? 8.

22. Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., and Kriegeskorte, N. (2014). A Toolbox for Representational Similarity Analysis. PLoS Comput Biol 10, e1003553. 10.1371/journal.pcbi.1003553.

23. Lockwood, P.L., Wittmann, M.K., Nili, H., Matsumoto-Ryan, M., Abdurahman, A., Cutler, J., Husain, M., and Apps, M.A.J. (2022). Distinct neural representations for prosocial and self-benefiting effort. Current Biology 32, 4172-4185.e7. 10.1016/j.cub.2022.08.010.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Reviewer #2 (Recommendations for the authors):

I thank the authors for their thoughtful and thorough responses to reviewer comments. My past review raised two key concerns. First, I suggested that the paper could engage more deeply with past work on the computational basis of social learning, in order to ground their discussion on the differences between social and non-social learning. Second, I raised the concern that condition differences between social and non-social stimuli could be driven by lower-level features, such as the attentional salience or visual distinctiveness of the stimuli. Overall, I believe that the authors have gone above and beyond to address these concerns.

We thank the reviewer for these kind words and appreciate that our effort to address all comments are acknowledged. We believe that the suggested changes that were raised by the reviewer have improved the manuscript – now the findings are better embedded in previous literature, while they are also additionally extended with a theoretical Bayesian model.

First, the authors revised their introduction and discussion to cite past scholarship in social learning. But, beyond that, they also substantially expanded their analyses to disentangle differences between the underlying mechanisms driving behavior in social and non-social conditions. The authors found that participants did not have different prior expectations about the performance of social vs. non-social predictors. Instead, using a modified, noisy Bayesian model, the authors suggest that differences between conditions may be driven by degrees of uncertainty in the belief update and that greater noise in the update process leads to larger changes in confidence across different encounters with the predictor. I agree that these analyses have significantly improved the manuscript and made a new theoretical contribution in their own right, as they suggest a common computational mechanism underlying observed behavioral differences between conditions.

We thank the reviewer again for acknowledging our work and the effort we have put into the revision. We are happy to hear that the reviewer also believes that these additional analyses have extended our prior results.

If I can offer a small suggestion: Because these analyses and hypotheses were not planned in the initial submission, it would be helpful to specify that they are exploratory/an extension when they are first presented in the main text (p. 10-11).

We have made the following changes to the paragraph related to the Bayesian analyses, as well as to the discussion:

Paragraph: ‘The impact of noise in belief updating in social and non-social conditions’, page 10:

“In further exploratory analyses, we used Bayesian simulation analyses to investigate whether a common mechanism might underlie these behavioural differences. We tested whether the integration of new evidence differed between social and non-social conditions; for example, recent observations might be weighted more strongly for non-social cues while past observations might be weighted more strongly for social cues.”

Discussion, page 15:

“An alternative explanation might be that participants experience a steeper decline of subjective uncertainty in their beliefs about the accuracy of social advice, resulting in a narrower prior distribution, during the next encounter with the same advisor. During a series of exploratory analyses, we used Bayesian simulations to investigate how uncertainty about beliefs changed from trial to trial and showed that belief updates about non-social cues were consistent with a noisier update process that diminished the impact of experiences over the longer term. From a Bayesian perspective, greater certainty about the value of advice means that contradictory evidence will need to be stronger to alter one’s beliefs.”

Second, the authors added several control analyses as a supplement. In order to test whether condition differences could be accounted for by attentional engagement, the authors inspected the constant of the GLM. They found no condition differences unaccounted for by the parametric regressors included in the GLM, either in attentional regions or in the regions that were the focus of their study. Next, in order to test whether social and non-social stimuli differed in their visual distinctiveness, the authors used RSA to compare the EDI between conditions in early visual areas. Again, the authors found no difference, suggesting that observed condition differences are unlikely to be driven by these lower-level features.

Thank you for providing such a comprehensive summary, we appreciate the effort that went into reviewing our manuscript.

Author details

1. Nadescha Trudel

   1. Wellcome Centre of Integrative Neuroimaging (WIN), Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom
   2. Wellcome Centre for Human Neuroimaging, University College London, London, United Kingdom
   3. Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, London, United Kingdom

   Contribution

   Conceptualization, Data curation, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing

   For correspondence

   nadeschatrudel@gmail.com

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-9372-3640

2. Patricia L Lockwood

   1. Centre for Human Brain Health, School of Psychology, University of Birmingham, Birmingham, United Kingdom
   2. Institute for Mental Health, School of Psychology, University of Birmingham, Birmingham, United Kingdom
   3. Centre for Developmental Science, School of Psychology, University of Birmingham, Birmingham, United Kingdom

   Contribution

   Resources, Software, Methodology, Writing – review and editing

   Competing interests

   No competing interests declared

3. Matthew FS Rushworth

   1. Wellcome Centre of Integrative Neuroimaging (WIN), Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom
   2. Wellcome Centre of Integrative Neuroimaging (WIN), Centre for Functional MRI of the Brain, Nuffield Department of Clinical Neurosciences, John Radcliffe Hospital, University of Oxford, Oxford, United Kingdom

   Contribution

   Conceptualization, Supervision, Funding acquisition, Methodology, Writing – review and editing, Investigation, Resources

   Contributed equally with

   Marco K Wittmann

   Competing interests

   No competing interests declared

4. Marco K Wittmann

   1. Wellcome Centre of Integrative Neuroimaging (WIN), Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom
   2. Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, London, United Kingdom
   3. Department of Experimental Psychology, University College London, London, United Kingdom

   Contribution

   Conceptualization, Supervision, Methodology, Writing – review and editing, Resources

   Contributed equally with

   Matthew FS Rushworth

   For correspondence

   m.wittmann@ucl.ac.uk

   Competing interests

   No competing interests declared

• 1,020

  Page views

• 130

  Downloads

• 0

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.