1. Introduction

Adolescents are generally more preoccupied with their peers and more prone to peer influence than in other periods of life (Andrews et al., 2021, Blakemore and Mills, 2014, Somerville, 2013). Peer influence, or social information use, can be very advantageous, such that adolescents can use their peers as sources of information on how to behave adaptively and morally within their changing environment (Bingham et al., 2016, Duell and Steinberg, 2019, Telzer et al., 2018, Molleman et al., 2022). Moreover, imitating others’ behavior can speed up learning, by not having to go through a costly process of trial-and-error (Morin et al., 2021).

However, not all peer information is equally valuable and in not every situation one should be distracted by peers’ opinions. Learning when to rely on peers is a crucial developmental goal for adolescents. The value of using social information is dependent on 1) our current knowledge, and 2) the expertise of the source. First, when we find ourselves in new or uncertain environments, and thus have little knowledge about what to do, we are more inclined to rely on others’ behavior to inform our own actions (Ciranka and van den Bos, 2020; Laland, 2004). In this respect, it might not be very surprising that adolescents are generally more prone to peer information, as they find themselves in a very volatile period in which they encounter many new and unstable (social) situations (Ciranka and van den Bos, 2021, Parr et al., 2024). Indeed, previous empirical studies have found more information search and social information use for adolescents compared to adults (Moutoussis et al., 2016, Niebaum et al., 2022). However, although both adults and adolescents (Slagter et al., 2023) use more social information when they are more uncertain, it is not yet clear how adolescents’ sensitivity to uncertainty compares to that of adults.

Apart from your own uncertainty, the expertise of the other poses an important factor in determining how to weight their information relative to your own. In daily life, however, objective measures of others’ expertise are often lacking. Instead, others often communicate how confident they are, with both children and adults being more likely to use social information from more confident peers (Gradassi et al., 2022a, Wood et al., 2013, Zarnoth and Sniezek, 1997). Although adolescents are generally more focused on their peers (Andrews et al., 2021, Blakemore and Mills, 2014, Somerville, 2013), and they do use social information depending on the peer’s intelligence and social status (da Silva Pinho et al., 2021; Gradassi et al., 2022b; Helms et al., 2014; Wilks et al., 2015), their response to peer confidence has not yet been fully explored.

Previous research strongly suggests social information use can be described as akin to Bayesian updating, where people engage in precision weighting of the various pieces of information (Behrens et al., 2008, Ciranka and van den Bos, 2020, Perreault et al., 2012, Toelch et al., 2014). This means that more certain or reliable information is considered to be more precise and therefore given greater consideration. Thus, when you are more certain about your personal information – that is, outcomes that you have experienced yourself – you will more strongly weight this information compared to social information (Fig. 1A). Conversely, when a peer reports to be very confident, you will translate this into a high precision weight and a stronger belief update based on their information. Apart from Bayesian principles, adolesecents and adults also apply simple heuristics to use social information, including copying peers if they constitute a majority or when they display consistent behavior (Kendal et al., 2018; Molleman et al., 2020; Slagter et al., 2023). For example, when adult participants played a task in which they could update their initial response based on social information, their behavior was best described by a model that included both Bayesian updating to combine personal and social information and a simple stay heuristic to stick with their initial response, without considering the peer’s response (Molleman et al., 2020). It remains unclear how Bayesian processes and heuristic strategies evolve throughout adolescence. While complex reasoning abilities tend to improve with age during this period (Huizenga et al., 2007), there is also evidence that mid-adolescents can exhibit more optimal Bayesian behavior in tasks requiring adaptation to changing environments (Eckstein et al., 2022). Similarly, some heuristics, such as a bias toward random exploration, appear to decrease with age (Dubois et al., 2022), though the direction of change may vary depending on the specific bias in question (Strough et al., 2011).

These developmental shifts in decision-making may be associated with ongoing neural changes. For example, in adults, uncertainty (or precision) has been shown to be represented across a wide network within the brain. Bayesian probability distributions can already be decoded from the sensory cortex (Geurts et al., 2022, van Bergen et al., 2015) Higher cognitive areas, such as the medial prefrontal cortex (mPFC) (Bach and Dolan, 2012, De Martino et al., 2013, Gherman and Philiastides, 2018, Lebreton et al., 2015, Rouault et al., 2023), the anterior cingulate cortex (ACC) (Bang and Fleming, 2018, Fleming et al., 2012), the anterior insula (Bach and Dolan, 2012, Singer et al., 2009) and the ventral striatum (Bang and Fleming, 2018, Hebart et al., 2016) have been implicated in subjective confidence, and the dorsolateral prefrontal cortex (dlPFC) has been implicated in outcome uncertainty and metacognition about uncertainty (Alexander et al., 2023, De Martino et al., 2013, Fleming et al., 2012, Rouault et al., 2023). Many of these areas, including the striatum, mPFC and dlPFC have also been identified in adults in reasoning about the value and uncertainty of others’ information, including peer confidence reports (Burke et al., 2010, Heyes, 2012, Ruff and Fehr, 2014). Other areas, including the temporoparietal junction (TPJ) (Campbell-Meiklejohn et al., 2017, McKell Carter et al., 2012, Saxe and Kanwisher, 2003, Zaki et al., 2016) and the superior temporal sulcus (STS) (Isik et al., 2017, Rushworth et al., 2013) are more specifically involved in social cognition and Theory of Mind (ToM), which involves reasoning about others' intentions and beliefs. For example, the TPJ has been shown to play a role in the tracking of social information (Zhang and Gläscher, 2020) and expertise (Boorman et al., 2013). Moreover, correlates of the integration of personal and social information have been observed in the ventromedial prefrontal cortex (vmPFC) and ventral anterior cingulate cortex (Behrens et al., 2008, Lockwood and Wittmann, 2018, Zhang and Gläscher, 2020).

Functional changes within these neural networks during adolescence can lead to differential processing of uncertainty and peer-related signals between adolescents and adults. Ongoing maturation of the PFC during adolescence (Casey, 2015, Casey et al., 2008, Gogtay et al., 2004) might result in poorer precision-estimates and less accurate forms of metacognition. The question then arises to what extent adolescents’ higher order decision-making resembles Bayesian updating. Not perfectly integrating precision-estimates into their decisions could lead adolescents to be more uncertain in general (Reiter et al., 2021) and to be less sensitive to varying levels of state uncertainty and peer confidence, rendering them more susceptible to peer influence and to using simple decision heuristics instead. The ongoing development of ToM and the TPJ during adolescence (Güroǧlu et al., 2011, Klapwijk et al., 2013, Mills et al., 2014), could additionally contribute to adolescents being less sensitive to peer-related cues. On the other hand, adolescents might be more sensitive to peer cues due to their strong focus on peers, which could be influenced by heightened activation of striatal regions (Blakemore and Robbins, 2012, Crone and Dahl, 2012, Somerville et al., 2011).

Here, we will compare if and how processes related to social information use under uncertainty differ between adolescents and adults. We will use a social information use task, in which participants have the opportunity to update an initial response based on a peer’s response. We vary both the amount of their personal information, affecting their own certainty, as well as the peer’s confidence rating. Based on previous research suggesting that social information use is subject to Bayesian information processing, we compare different Bayesian models to computationally describe participants’ cognitive processes in terms of a precision-modulation of both personal and social information, which is then translated into a Bayesian update. The Bayesian model is then further expanded to include a decision heuristic that captures participants' tendency to stay with their initial response, regardless of the peer's input. Our primary, hypothesis-driven predictions are that adolescents show higher overall social information use and reduced sensitivity to their own certainty. Our predictions for sensitivity to peer confidence are less clear: adolescents may be either less sensitive due to immature (meta)cognition or more sensitive due to heightened peer focus. We also explore neural activity during the task using fMRI to investigate how neural responses to personal and peer-related certainty differ between adolescents and adults. Given the previously discussed neural differences in decision-making between these age groups, we are particularly interested in examining activity within the PFC, insula, TPJ, and striatum.

2. Methods

2.1. Participants

One hundred four adult (22–45 years old) and 105 adolescent (12–18 years old) participants were recruited via the research participant system of the University of Amsterdam, Instagram or posters at schools in the vicinity of the research center. Participants were right-handed, Dutch-speaking, screened for MR compatibility, had normal or corrected-to-normal vision and had completed or were following a university-level education (adults) or had completed or were following pre-university-level education (adolescents). The experiment described in the current study was part of a larger task battery, consisting of a session from home (30 min) during which participants filled in a questionnaire about their digital device and social media use and a behavioral-fMRI session at the research center (2.5 h) during which they completed a social information use task in the MR scanner (described below) as well as a spatial working memory task, a reversal learning task, an information search task and a shorter social information use task. This latter task, the Berlin Estimate Adjustment Task (Molleman et al., 2019, 2020), was performed approximately 30 min after completing the task in the scanner. Similar to the task described in the current manuscript, participants estimated the number of animals shown on screen, viewed the estimate of an unknown peer, and had the opportunity to revise their response. Participants were paid 40 euro plus a bonus contingent on their performance, ranging between 0.70 and 2.90 euro for all tasks combined, and ranging between 0 and 1.50 for the fMRI task. All participants performed the social information use task inside the MR scanner, and both behavioral and fMRI analyses were conducted on data from this single task. Out of the participants who signed up for the study, a total of 74 adults and 70 adolescents completed their fMRI session. From the adult group, 2 were excluded due to incomplete datasets, one was excluded because they fell outside our predefined age range, 3 participants were excluded based on our behavioral exclusion criteria described below, and a further 10 were excluded from the fMRI analysis because they did not pass our quality assessment (9) or because their timing files were unavailable (1), resulting in 68 participants included in the behavioral analysis (27 men, 40 women, and 1 participant who identified as non-binary; age: mean (SD) = 24.8 (3.7); Fig. 1D) and 58 participants included in the fMRI analysis (22 men, 35 women, and 1 participant who identified as non-binary; age: mean (SD) = 24.3 (3.5)). From the adolescent group, 6 participants were excluded based on our behavioral exclusion criteria and a further 17 were excluded from the fMRI analysis because they did not pass our quality assessment, resulting in 64 participants included in the behavioral analysis (23 men, 39 women, and 2 participant who identified as non-binary; age: mean (SD) = 16.4 (1.5)) and 47 participants included in the fMRI analysis (18 men, 27 women, and 2 participant who identified as non-binary; age: mean (SD) = 16.4 (1.5); Fig. 1D). The procedure of both studies was approved by the local ethics committee (Adults: 2021-DP-13756; Adolescents: 2021-DP-13738). Before the start of the experiment, participants, and their parents in case the participant was younger than 16 years old, gave written consent according to the declaration of Helsinki.

2.2. Social information use task

The social information use task is an estimation game modelled after the marble task (Phillips and Edwards, 1966) and the Berlin Estimate AdjuStment Task (BEAST; (Molleman et al., 2019)) and probes the amount of social information use as a function of own and others decision confidence. The game was programmed in Python 3.6.6 using PsychoPy. A schematic of the game and corresponding timings is schematically depicted in Fig. 1B.

Participants were instructed that they are part of a community for which the main food resource is mushrooms. These mushrooms are harvested from a forest in the vicinity of their village. Soon, it will be time to harvest, and the leader of the community has asked them for their help in coordinating this. There are two types of mushrooms, blue and red ones. Their task at the start of each day (trial) is to estimate the percentage of blue mushrooms in the forest. To make an informed guess, they will walk through 5 fields that surround their village. The mushrooms in the fields are a good, but not exact representation of what will be found in the forest on that day. On some days they encounter many mushrooms (9 mushrooms per field, 45 in total) in the fields and on other days they encounter fewer mushrooms (1 mushrooms per field, 5 in total). The more mushrooms they find in the nearby fields, the more information they have to base their estimate on and the more certain they can be. Participants submitted their initial estimate using a 100-point slider (ranging from 0 to 100) operated via an MR-compatible button box. This estimate reflected their belief about the proportion of blue mushrooms, with the red mushroom value automatically set so the total equaled 100. After submitting this initial estimate, participants were shown peer information stemming from a peer described as a previous participant in the study. No additional information about the peer was provided. The percentage of blue and red mushrooms in the forest was the same for this other person, but they were from a different village on the other side of the forest, meaning that they had to sample from different fields on the other side of the forest, and the number of blue and red mushrooms in these fields might have been different from those in the fields surrounding the participants village. The participant was first shown the peer’s confidence rating in their estimate, which could be either low, medium or high. After 4 s, this information was accompanied by the peer’s estimate. Based on this new information, the participant could adapt their estimate. The peer’s estimate and confidence rating stemmed from real previous participants, who had played a version of the mushroom game without any social information. The peer estimate was chosen such that, if possible, it was in the direction of the true underlying percentage and the difference between the participants first estimate and the peer estimate ranged between 15 and 18 percentage points. This range was selected to leave enough room for adjustment of the estimate and control for distance that might influence the amount of adjustment (Gradassi et al., 2022a, Gradassi et al., 2022b, Molleman et al., 2019, Molleman et al., 2020, da Silva Pinho et al., 2021).

Participants completed 75 trials, pseudo-randomly divided across 3 runs, which lasted approximately 30 min in total. Sixty trials were equally divided across own certainty (certain, uncertain), peer confidence (low, medium, high) and true percentage of blue versus red mushrooms (12.5, 25, 37.5, 62.5, 75). Another set of 15 “filler trials” was included to make the trials more realistic: the true percentage could range from 0 to 100 and the peer’s estimate was either very close (< 3) or very far (> 40) from the participant’s first estimate. The maximum time to submit an estimate was 20 s, after which the trial would time out and the participant would see a message stating that they were too late. No feedback was given on their estimates, but participants could earn a bonus depending on their performance. One estimate of one trial was randomly selected. If this estimate was exactly the same as the true percentage, they would earn 1.50 pounds, of which 0.04 pounds was subtracted for each percentage point off. Their bonus could never go below 0.

The main task was preceded by instructions, 2 short rounds of 3 practice trials each, and a longer practice round of 15 trials outside the scanner. During this practice round, participants were asked to rate their own confidence after submitting their first estimate as either low, medium or high, to give them a better understanding of what the peers confidence ratings entailed. Moreover, a quick analysis of this practice round informed the experimenter about whether the participant indeed understood the game before going into the scanner. Specifically, if a participant’s initial estimates followed the absolute number of blue mushrooms rather than the ratio of blue to red mushrooms, or if their estimates showed no sensitivity to the ratio, the task was explained again. The participant was then asked to explain the task back to the experimenter to ensure they had understood it correctly.

2.3. MRI acquisition and preprocessing

The fMRI experiment was performed on a Philips Achieva 3 T MRI scanner and a 32-channel SENSE headcoil at the imaging center at the Roeterseiland campus of the University of Amsterdam. First, a 6-minute whole-brain high-resolution structural image was acquired for anatomical referencing, using a T1-weighted magnetization prepared, rapid-acquisition gradient echo (MPRAGE) sequence (TR = 8.2 ms, TE = 3.7 ms, flip angle = 8°, 220 axial slices in ascending order, field of view = 240 × 188 mm, voxel size = 1 × 1 × 1 mm). Images with blood-oxygen level-dependent (BOLD) contrast were acquired in 3 runs, using a whole-brain T2*-weighted single-shot gradient-echo multi-echo echo planar imaging (EPI) sequence (TR = 2000 ms, TE = 28 ms, flip angle = 76.1°, 36 axial slices per volume in ascending order, field of view = 240 × 240, reconstruction matrix = 80 × 80, voxel size = 3 × 3 × 3 mm, slice gap = 0.3 mm). After each functional run, a phase-encoding polarity (“topup”) gradient-echo sequence consisting of 3 volumes was acquired to estimate the static magnetic field’s inhomogeneity. The parameters were exactly the same as for the corresponding functional scans, but with a reversed phase-encoding direction.

All images were converted from PAR/REC to NIfTI format and into BIDS (Brain Imaging Data Structure) (Gorgolewski et al., 2016) using BIDScoin (Zwiers et al., 2022). The images were preprocessed using HALFpipe (Harmonized AnaLysis of Functional MRI pipeline) version 1.2.2 (Waller et al., 2022), which is based on fMRIPrep (Esteban et al., 2019). Structural images were corrected for bias field, skullstripped and segmented before being registered to the MNI152NLin2009cAsym standard template. Preprocessing of the functional images included realignment to the middle volume, slice timing correction, susceptibility distortion correction based on the topup scans, coregistration to the participant’s structural scan and spatial normalization to MNI152NLin2009cAsym space. Denoising was then performed by spatial smoothing using a 6 mm FWHM kernel, grand mean intensity normalization with a mean of 10 000 and temporal filtering using a high pass filter width of 100 s. Physiological nuisance regressors were extracted using aCompCor (Behzadi et al., 2007). See https://fmriprep.readthedocs.io/en/0.6.2/workflows.html for more details on the preprocessing pipeline. Exclusion after quality assessment was based on a visual check of the skull stripping, spatial normalization, EPI tSNR, EPI confound and head motion results (excessive motion was defined as mean FD > 0.5 mm or max FD > 3 mm).

2.4. Behavioral data analysis

Our main outcome measure was social information use (s), defined as the adjustment from the first (E₁) to the second (E₂) estimate, relative to the peer estimate (P):(1)

A value of 0 indicates no social information use, 1 indicates complete copying of the peer, negative values reflect adjustment away from the peer, and values above 1 indicate over-adjustment beyond the peer estimate. Participants who did not understand the task, as assessed per qualitative visual inspection, were excluded from the analysis (e.g. participants who responded randomly to differences in percentage of blue versus red mushrooms, or participants who did not extrapolate the number of seen mushrooms to a total of 100), as were participants whose social information use was 0 at more than 70 % of the trials, as this would yield minimal variation to conduct the analyses. Filler trials, missed trials (without a response at either the first or second estimation phase), trials in which no suitable social information (i.e. within the correct range from the participant’s first estimate) could be sampled from our database, trials for which either the first or the second estimate deviated more than three standard deviations from the group mean, grouped per own certainty condition, ratio and age group, and trials in which social information use did not range between 0 and 1, meaning that it was not a weighted average of the participant’s first estimate and the peer’s estimate,¹ were excluded from the analysis (Table S1).

The behavioral analyses were conducted in R version 4.0.4 (R Core Team, 2021). To assess the combined effects of own certainty and peer confidence on social information use, we conducted a linear mixed effects regression, using the lmer function from the lmerTest package (Kuznetsova et al., 2017). Own certainty and peer confidence (with mean-centered values −1 for low, 0 for medium and 1 for high peer confidence), as well as their interaction, were included as both fixed and random effects:

To compare age groups, we ran the same analysis again while including the categorical variable group (adolescents or adults) as a fixed effect, both as a main effect and in interaction with the other predictors:

Intraclass correlation coefficients (ICCs) were computed using the performance package (Lüdecke et al., 2021) to quantify the proportion of variance attributable to between-subject differences, and reported marginal and conditional R² to quantify the variance explained by the fixed effects alone and by the full model, respectively. Standardized beta coefficients (β) were computed using the standardize_parameters() function from the parameters package (Lüdecke et al., 2020).

2.5. Control analyses

To verify that any group differences in social information use could not be attributed to differences in performance, we ran a linear mixed effects regression on performance. Performance was indexed by deviance, defined as the absolute difference between the participant’s first estimate and the true percentage of blue mushrooms. The model included age group and own certainty as fixed effects, along with their interaction. To account for within-subject variability, we included random intercepts and random slopes for own certainty at the participant level:

To further explore whether differences in performance, and possibly in participants’ subjective confidence, could underlie age differences in the effect of own certainty, we conducted an additional analysis using confidence ratings collected during the practice phase. Specifically, we tested whether adolescents differ from adults in their ability to align confidence with performance. For each participant, we computed the correlation between deviance and confidence ratings, after which we compared the resulting subject-specific R² values across age groups using a t-test.

To test whether potential age group differences in residual variance affected our results, we conducted a control analysis using the lme function from the nlme package (Pinheiro et al., 2025), allowing for different residual variances by age group. See Supplementary Section Modeling Heteroscedasticity for more details.

In the above analyses, we treated age as a categorical variable (adolescents vs adults), which we considered the most appropriate approach given the relatively discrete age clusters in our sample (16–18 vs 22–23 years; see Fig. 1D). This group-wise approach allowed for clear hypothesis testing regarding age-related differences in social information use and neural mechanisms. However, to assess the robustness of our findings, we also conducted supplementary analyses using age as a continuous predictor (see Supplementary Section Age as a continuous predictor).

2.6. Computational modeling

We used a computational modeling approach to quantify Bayesian processes that we hypothesized to underly participants’ choices (Fig. 1A). Apart from a base model M0, we developed nine models with increasing complexity to investigate whether additional parameters explained the observed behavior better. The model-building process was conditional: If a model performed worse compared to the preceding model (based on the Bayesian Information Criterion; see Model fitting procedure below), we did not use it as the basis for further model development. Instead, the next model in the sequence was built upon the last accepted model. This iterative approach ensured that each new model only added complexity if it improved the model fit. Parameter bounds were chosen based on pilot data (Gershman, 2016) to avoid edge effects in posterior distributions while allowing a wide range of plausible behaviors (e.g., α and θ parameters could produce subjective mushroom counts well beyond the actual stimuli, up to 4500 for self-estimates in the certain condition and 1500 for peer estimates under high peer confidence).

An overview of all models can be found in Table 1. For all models, we assumed a beta distribution over all possible response options for their E2 (0–100 %, divided by 100). We set an initial prior of 1 blue and 1 red mushroom, to which we then added the participant’s estimate of blue and red mushrooms to arrive at a posterior for E1, where we further assumed that participants’ E1 was an accurate representation of the percentage of blue mushrooms they had seen in the fields:(2)where is the total number of mushrooms that the participant saw on that trial, being either 5 for the uncertain condition and 45 for the certain condition. For example, if 45 mushrooms were shown in a particular trial and the E1 was 20, meaning that the participant estimated that the percentage of blue mushrooms was 20 %, the number of blue mushrooms was set to 20/100 × 45, which was then added to the prior of 1, resulting in . The number of blue and red mushrooms as seen by the peer follow from the peer’s estimate P and the total number of mushrooms seen by the peer. Because the total number of mushrooms seen by the peer is unknown to the participant, we set this to 25 in M0, which is the average of the total number seen by the participant, across all trials.

Adding these peer’s and to the participant’s and , respectively, results in the posterior for E2:(4)

One can then define a probability density function (PDF) of the posterior beta distribution based on and , for and shape parameters and :(5)where the beta function is a normalization constant to ensure that the total probability adds up to 1, which we then used for our model fitting procedure (see more details below).

2.7. Model family 1: Individual modulation of the weight of E1

Participants might under- or overestimate the numerosity of displayed items or their own confidence (Krueger, 1984, Lichtenstein and Fischhoff, 1977). In M1a, we therefore tested whether adding a parameter that modulates the perceived number of total mushrooms, effectively modulating the subjective weight of the E1 at the individual level, improved model performance:(6)

In M1b, we extended this modulation by including a dual , which separately modulated the perceived number of total mushrooms for the uncertain and certain conditions:(7)

2.8. Model family 2: Individual modulation of the weight of the peer estimate

Next, we tested whether individually modulating the number of mushrooms the participant assumed the peer had seen improved model performance. In M2a, rather than fixing the number of mushrooms as seen by the peer to 25, we introduced that modulated this number per participant:(8)

In M2b, we tested whether model performance improved when the modulation by linearly depended on the confidence level reported by the peer, such that the weight of the peer estimate increased as the peer reported to be more confident:(9)where θ_IC is the intercept representing the baseline influence of the peer when confidence is at its lowest level, and θ_slope captures how much the influence increases with each step in confidence. Peer confidence level was defined as 1 for low, 2 for medium and 3 for high confidence. In M2c, we tested whether a modulation by for each peer confidence level separately, allowing for a non-linear modulation, improved model fit:(10)

2.9. Model family 3: stay bias

A visual inspection of the data made clear that participants exhibited a tendency to stay with their first estimate (s = 0), a response pattern that stood apart from the rest of their near-normally distributed response pattern. This suggests that at least at some trials, participants exhibited a bias in staying with their first estimate, without even processing the peer’s estimate, in line with earlier findings suggesting that people combine Bayesian processing with simple heuristics. To allow for this bias, we added a stay bias to models M3. For M3a the stay bias, defined by parameter , was non-selective, such that it was set per individual, but equal across all trials within-participant. M3b included a separate parameter β for each own certainty condition, and . For M3c and M3d, the stay bias depended on peer confidence level in a linear or exponential fashion, respectively:(11)

For these M3 models, the PDF was then multiplied by , such that the sum of the stay bias and the total area under the curve (the probability) of the PDF adds to 1. The probability of a stay response at E2 was then estimated by the sum of the stay bias and the probability of s = 0 under the PDF. This means that a stay response could be either the result of a simple stay bias, or of a decision process that considered the peer’s estimate. On non-stay trials, the probability of participants’ responses was estimated solely by the probability of the E2 response under the PDF (multiplied by ).

To summarize, this last model combines Bayesian integration of personal and social information with a stay bias component. Participant’s own evidence is transformed into subjective mushroom counts:where E1 is the participant’s first estimate (0−100), N is the total number of mushrooms shown and α modulates certainty differentially for each of the two own certainty conditions.

Next, the peer’s estimate P is transformed into subjective mushroom counts linearly based on the peer’s confidence rating using a θ intercept and slope:

The posterior belief is then computed as a Beta distribution, which defines a probability density over all possible values of social information use ranging from 0 (fully relying on one's own estimate) to 1 (fully copying the peer). with shape parameters:

Finally, the model includes a stay bias β to account for ignoring the peer information entirely. This results in a mixture distribution that combines a Beta distribution with a delta function at zero social information use:

2.10. Model fitting procedure

Parameter estimations were performed by individually fitting the model predictions to each participant’s behavior, specifically their E2 responses. The fitting procedure relied on a combination of grid search and maximum likelihood estimation using the L-BFGS-B algorithm implemented in the optim function in R. The log likelihood of the combination of parameter values corresponding to each point on the grid was estimated and minimized. Model comparison was evaluated by computing the Bayesian Information Criterion (BIC), which penalizes for increasing model complexity, across all participants for each of the models. Lower BIC values indicate a better fit. To evaluate differences in cognitive processes between adolescents and adults, we subsequently tested each parameter for statistical differences between age groups using Mann-Whitney U test. In order to account for multiple comparisons, we applied the Benjamini-Hochberg (BH) correction to control the false discovery rate.

As for our model-free analyses, we also conducted supplementary analyses using age as a continuous predictor (see Supplementary Section Age as a continuous predictor).

2.11. Model recovery and parameter recovery

For quality control, we ran a model recovery procedure using nine models (excluding base model M0). For each model, we simulated a dataset consisting of 132 simulated participants (68 corresponding to the Adult group and 64 corresponding to the Adolescent group). To simulate realistic parameter values, we sampled from a uniform distribution bounded by the minimum and maximum values observed in the real data after excluding outliers beyond ±3 standard deviations. Each of these simulated datasets were then fit by all nine models – resulting in 84 model fitting procedures. These fits were compared regarding their goodness-of-fit (BIC) and a confusion matrix was constructed to quantify model recoverability (Wilson and Collins, 2019).

We further ran a parameter recovery procedure on all our models, where we assessed the correlation between the ‘true’ parameter estimates, which we defined as the simulated parameter values, and the fitted parameter estimates, which were generated by rerunning each model on simulated data based on these ‘true’ estimates in combination with the same settings and trials as in the experiment.

2.12. fMRI data analysis

FMRI data processing was carried out using FEAT (FMRI Expert Analysis Tool) Version 6.00, part of FSL (FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl ). Time-series statistical analysis was carried out using FILM with local autocorrelation correction (Woolrich et al., 2001). For each run of each subject, BOLD responses were modeled using event-related GLMs. The first GLM included two parametric regressors: one that modeled model-based own certainty at the time of the E1 (i.e. the prior; first 500 ms, so brain activity would reflect (un)certainty rather than motor activity; Fig. 1B) and one that modeled model-based peer confidence at the time of the social information (duration of 4000 ms). Model-based own certainty was the actual number of mushrooms that was shown on that trial multiplied by either or (see Eq. 7), depending on the trial condition. Model-based peer confidence was defined as with peer confidence set as low = 1, medium = 2, high = 3 (see Eq. 9). We reasoned that model-based own certainty and peer confidence, based on the parameter values resulting from our best fitting Bayesian model, yielded more individualized results. However, we also ran a supplemental GLM that included two regressors relating to the model-free conditions of own certainty, modeled at the time of the E1 (first 500 ms), one for the certain and one for the uncertain condition, and one regressor relating to model-free peer confidence modeled at the time of the social information (duration of 4000 ms). The latter was parametrically modulated by the peer confidence level. We created contrasts for the effect of own certainty (certain – uncertain at E1) and the effect of peer confidence (parametric effect at social information).

The parametric regressors were mean-centered across all trials within a run. Of note, peer confidence, which was modeled at the time of social information, only included non-filler trials, because filler trials contained peer information (very close or very far from the E1) that might have elicited different cognitive processes. In both models, we additionally included temporal derivatives, six motion parameters, framewise displacement, global white matter and CSF signal and the first 6 anatomical principal component noise regressors (aCompCor) as regressors of no interest. All regressors were convolved with a double-gamma hemodynamic response function.

The resulting contrast images per run were averaged at the participant level, which was carried out using a fixed effects model, by forcing the random effects variance to zero in FLAME (FMRIB's Local Analysis of Mixed Effects) (Beckmann et al., 2003, Woolrich, 2008, Woolrich et al., 2004). The participant-level contrasts were subsequently taken to the group level, carried out using FLAME stage 1 to account for both within- and between subject variability. We first included all participants in a combined analysis to identify brain regions activated across the entire sample using a one-sample t-test. We then performed a direct comparison between adolescents and adults using two-sample t-tests to identify clusters showing significant differences in brain activation between the two groups. To further explore the clusters identified in the group comparison, we conducted separate follow-up analyses for adolescents and adults, each using a one-sample t-test. This allowed us to determine whether the clusters identified in the group comparison exhibited significant activation within each group individually. Normalized (Z-scored) statistic images were thresholded using clusters determined by Z > 3.1 and a (corrected) cluster significance threshold of P = 0.05 (Worsley, 2001). For the group-level analyses, we applied a mask that included the frontal lobe and insula, derived using the MNI structural atlas in FSLeyes, as well as the TPJ and the striatum, based on connectivity-analyses (Mars et al., 2012, Piray et al., 2017) (Fig. 1C). Using these predefined regions provided practical and standardized anatomical boundaries while ensuring complete coverage of the areas of interest described above. However, to also give a more complete overview, results pertaining to age differences at the level of the whole-brain can be found in the supplemental material (Table S8-9).

3. Results

3.1. Effect of own certainty and peer confidence on social information use

For both groups, participants first estimates (E1) clearly followed the true percentages, indicating that they understood the game (Fig. 2A). The average social information use across all trials was 0.37 (Fig. 2B,C), which is comparable to studies using similar designs (Gradassi et al., 2022a, Molleman et al., 2019, Yaniv, 2004). This is consistent with an egocentric bias, where participants place more weight on their personal information than on the peer’s: a one-sample t-test confirmed that average social information use was significantly lower than 0.5 (t(131) = -11.52, p < 0.001). Our linear mixed effect regression model yielded an adjusted intraclass correlation coefficient (ICC) of 0.505, indicating that approximately 50 % of the variance was attributable to between-subject differences. The fixed effects alone explained 29 % of the variance (marginal R² = 0.286), while the combined fixed and random effects explained 65 % of the variance (conditional R² = 0.647). Across groups, social information use was higher in the uncertain compared to the certain condition (b = −0.12, β = −0.22, p < 0.001; Fig. 2B,C) and higher with increasing peer confidence (b = 0.16, β = 0.49, p < 0.001). There was also an interaction between own certainty and peer confidence (b = −0.02, β = −0.03, p = 0.009), such that the effect of peer confidence was stronger under uncertainty compared to certainty. The data thus indicate that participants displayed an egocentric bias, where they assigned roughly twice as much importance to their personal information compared to the peer's information. Only when they were uncertain themselves and the peer reported to be highly confident, they assigned equal or more weight to the peer information than to their personal information.

3.2. Age-related differences in sensitivity to own certainty and peer confidence

Our linear mixed effect regression model including age group as a regressor yielded an adjusted intraclass correlation coefficient (ICC) of 0.499, indicating that approximately 50 % of the variance was attributable to between-subject differences. The fixed effects alone explained 30 % of the variance (marginal R² = 0.297), while the combined fixed and random effects explained 65 % of the variance (conditional R² = 0.648). Adding age group as a fixed effect to the analysis showed no main effect of age group on social information use (M adults = 0.39, M adolescents = 0.36, b = −0.03, β = −0.06, p = 0.131), which is in contrast with earlier studies finding more social information use in adolescence (Andrews et al., 2021, Somerville, 2013). Notably, when our participants performed a similar task, albeit without any certainty manipulations, approximately half an hour later, adolescents relied significantly more on social information than adults (M adults = 0.31, M adolescents = 0.38, b = 0.07, β = 0.17, p = 0.003, Figure S1). This task difference emerged despite a positive correlation in average social information use across the two tasks (r = 0.23, p = 0.010). This may suggest that adolescents typically use more social information, but the presence of certainty and confidence cues in our task may have reduced this tendency. Crucially, age group did modulate the extent to which own certainty affected social information use (b = 0.09, β = 0.08, p = 0.001), such that adolescents were less sensitive to own certainty than adults (Fig. 2B,C). Age group did not significantly modulate the effect of peer confidence on social information use (b = −0.02, β = −0.03, p = 0.219), nor was there a significant three-way interaction between own certainty, peer confidence and age group (b = 0.02, β = 0.02, p = 0.102). To further explore the interaction between age group and own certainty, we used estimated marginal means to examine age-related differences at each level of own certainty. This revealed that the effect was driven by the uncertain condition, such that adolescents used less social information than adults in the uncertain condition (t(130) = 2.50, p = 0.014) but not in the certain condition (t(130) = -0.48, p = 0.635). Although adolescents were overall less sensitive to own certainty than adults, they still showed a significant effect of own certainty on social information use (adolescents: t(130) = 3.82, p < 0.001; adults: t(130) = 8.65, p < 0.001). An exploratory analysis on the participants’ mean confidence ratings, which they reported after each trial during the practice round (1 = low, 2 = medium, 3 = high confidence), revealed a similar effect. The difference between reported confidence in the uncertain versus the certain condition was smaller for adolescents compared to adults (b = −0.473, p < 0.001; Fig. 2D). Estimated marginal means revealed that this was mainly driven by adolescents feeling more confident than adults in the uncertain (adolescents: 1.99; adults: 1.60; t(258) = -4.30, p < 0.001) but not the certain condition (adolescents: 2.20; adults: 2.28; t(258) = 0.89, p = 0.377). To ensure group age differences in social information use were not driven by group differences in performance, we examined performance, indexed as deviance of participant’s first estimate from the true percentage of blue mushrooms as a function of group. This did not reveal a significant difference in performance between age groups (b = 0.04, p = 0.888; Fig. 2E), nor did performance depend on own certainty (b = −0.08, p = 0.715) or the interaction between group and own certainty (b = −0.05, p = 0.919). We next tested whether adolescents differ from adults in their ability to align confidence with performance. However, there was no significant group difference in the correlation between deviance and confidence ratings (R² adolescents = 0.119, R² adults = 0.118, t = 0.03, p = 0.975, Fig. 2F). Together, this suggests that age-related differences in the effect of own certainty on social information use are unlikely to be explained by differences in performance or in the mapping of performance to subjective confidence.

The effects reported above are based on a linear mixed-effects model that assumes homoscedastic residuals, or equivalent noise across age groups and conditions. As a control, we also tested whether allowing for age-group-specific residual variances would change the results. The heteroscedastic model showed improved model fit, but the fixed effects remained consistent, suggesting that the findings are robust to differences between age groups in noisy responding (Supplementary Section Modeling Heteroscedasticity).

3.3. Computational mechanisms of social information use

We used a computational modeling approach to quantify mechanisms hypothesized to govern social information use. Model comparison showed that, for both age groups, the model that best captured the data was M3d (Fig. 3A,B; Table 1), which included a dual α to modulate the impact of own certainty on the weight of the personal information at E1, a and that together modulated the weight of the peer estimate, linearly dependent on peer confidence, and a stay bias β to capture participants tendency to stay with their initial estimate, which became exponentially smaller as peer confidence increased:

Fig. 3C and D illustrate the parameter estimates and model predictions per age group. The dual α parameter allowed for a differential conversion of the actual number of mushrooms in each condition into a subjective count, reflecting differential translation into model-based certainty. Participants amplified their model-based certainty more in the uncertain condition (median = 29.4) compared to the certain condition (median = 5.3). When we multiply participants’ and with the actual number of mushrooms, we see that their model-based certainty was significantly lower in the uncertain (median = 146.8) versus the certain condition (median = 239.3; V = 1178, p < 0.001; Fig. 4 bottom right panel), such that participants behaved as if they were almost twice as certain in the certain versus the uncertain condition, even though the actual ratio between 45 and 5 is 9. Participants thus discounted the condition difference, mainly driven by a relatively exaggerated model-based certainty in the uncertain condition.

Further improvement of the model fit due to the addition of the parameters provides clear evidence of participants behaving as if the peer had seen more mushrooms for higher reported peer confidence. As per Eq. 9, a median of 54.8 and of 79.0 resulted in participants behaving as if the peer had seen 54.8 mushrooms for low confidence, 133.9 mushrooms for median confidence and 212.9 mushrooms for high confidence. Consequently, higher peer confidence led to a stronger weighting of the peer’s estimate in the participant’s second and final estimate.

Finally, better model performance after adding a stay bias that depended on peer confidence, but not on own certainty, shows that peer confidence is the driver of this decision heuristic. This indicates that even though both own certainty and peer confidence affected the extent to which participants used social information, changes of mind were more likely when peer confidence was high, but was not different for low versus high own certainty.

3.4. Model recovery and parameter recovery

The model recovery results show that data generated by one specific model generally also resulted in the best model fit for that same model, indicating successful model recovery (Figure S2). Importantly, simulations were not falsely fit best by our winning model M3d, suggesting that M3d does not overfit data generated by other models. However, there was some overlap in model fits between M3c and M3d, which is unsurprising given their strong similarity (a stay bias that depends linearly versus exponentially on peer confidence level.

Our parameter recovery procedure on M3d revealed high Pearson’s correlations between the ‘true’ and fitted parameter estimates (all r ≥ 0.60, all p < 0.001; Table S2) suggesting that the model was adequately able to recover the parameters (see Table S3 for correlations between the different parameters).

3.5. Age differences in computational processes

To evaluate differences in cognitive processes between adolescents and adults, we subsequently tested each parameter for statistical differences using a Mann-Whitney U test (Fig. 4). This revealed no significant difference in α in the uncertain condition (median adolescents = 28.4; median adults = 29.5; W = 2095, p = 0.714, p_corrected = 0.769), but a significant lower in the certain condition for adolescents compared to adults (median adolescents = 4.9; median adults = 5.9; W = 1629, p = 0.013, p_corrected = 0.032). This indicates that compared to adults, adolescents experienced a similar model-based certainty in the uncertain condition (adolescents: 28.4 * 5 = 142; adults: 29.5 * 5 = 148), but a lower model-based certainty in the certain condition (adolescents: 4.9 * 45 = 221; adults: 5.9 * 45 = 266). Consequently, this led to lower model-based difference in certainty between the conditions for adolescents compared to adults (W = 1482, p = 0.002; Fig. 4 bottom right). This is opposite to the pattern we found in the model-free analysis, where we saw that the group difference in sensitivity to own certainty was mainly driven by higher reported confidence and lower social information use in the uncertain condition for adolescents compared to adults.

Regarding peer confidence, there was no significant group difference in (median adolescents = 54.2; median adults = 55.3; W = 2111, p = 0.769, p_corrected = 0.769), indicating that adolescents and adults did not differ in how they interpreted low peer confidence. However, there was a significant group difference in (median adolescents = 62.0; median adults = 89.2; W = 1604.5, p = 0.009, p_corrected = 0.032). This indicates that with each increase in peer confidence level, adolescents behaved as if the peer had seen 62 more mushrooms, whereas adults behaved as if the peer had seen 89 more mushrooms. Thus, even though both groups allocated more weight to the peer estimate when peer confidence increased, compared to adults, adolescents allocated less weight to peer estimates when peer confidence was medium and high. This age-related difference was not present in the model-free results, highlighting the added value of the computational analyses in capturing subtle differences in how adolescents discount peer estimates more than adults, primarily at higher confidence levels.

There was no significant group difference in the value for the stay bias (median adolescents = 0.22; median adults: 0.15; W = 2527.5, p = 0.103, p_corrected = 0.172), suggesting that adolescents and adults were equally likely to change their minds after seeing the peer estimate.

3.6. Neural activity in response to model-based own certainty and peer confidence across age groups

First, we assessed neural activity at the time of the first estimate as a function of model-based own certainty (Fig. 5A, Table S4). There was a large network of regions that showed a positive correlation with certainty (certain > uncertain), including the dorsomedial PFC (dmPFC) and the bilateral dlPFC, anterior insula and left caudate nucleus. A different network showed a negative correlation with certainty (uncertain > certain), including the supplementary motor area, the mPFC, the bilateral Rolandic operculum, adjacent to the posterior insula, and areas within the bilateral TPJ. Results from our supplemental GLM that included model-free regressors of own certainty and peer confidence rendered results comparable to the individually-modeled regressors, validating the neural underpinnings of our computational model (Table S4; Figure S3C).

Next, we looked at neural activity at the time when the social information was presented as a function of model-based peer confidence (Fig. 5B, Table S5). Neural activity in the left precentral gyrus and caudate nucleus correlated positively with peer confidence, whereas a small cluster in the right middle frontal gyrus correlated negatively with peer confidence. Again, we found similar clusters for model-free peer confidence (Table S5; Figure S3D).

3.7. Age differences in neural activity in response to model-based own certainty and peer confidence

When comparing neural activity in response to model-based own certainty in adolescents versus adults at the time of the first estimate, we found that adolescents showed a stronger negative correlation in the anterior cingulate cortex and mPFC, the right Rolandic operculum and supplemental motor area (Fig. 6A; Table S6, also for model-free results). Indeed, post hoc analyses revealed that when analyzing the groups separately, we found a strong negative correlation of own certainty in these regions for adolescents and less so or not significantly for adults. This suggests that adolescents were more sensitive to own certainty than adults at the neural level, which is in contrast with the behavioral and computational results, which showed lower sensitivity to own certainty in adolescents. We therefore speculated that this higher neural sensitivity for adolescents at the time of the first estimate E1 might well have disappeared at the time of the social information, when personal and social information are being combined, such that at the time of social information, adults would show greater sensitivity to own certainty than adolescents. To investigate this, we ran the same model-based and model-free fMRI analyses again, but now modeled own certainty at the time of the social information rather than at E1. This revealed no differences in neural activity between adolescents and adults for either ways of modeling own certainty. Thus, whereas there was an age difference in sensitivity to own certainty, this was no longer significant at the time of integrating personal and social information.

When comparing neural activity in response to model-based peer confidence in adolescents versus adults at the time of social information, we observed significant differences in neural activity within the anterior and ventral parts of the mPFC and the mid-cingulate cortex for adults compared to adolescents (Fig. 6B; Table S7, also for model-free results). Analyzing the groups separately revealed a significant positive correlation with peer confidence in the mPFC for adults but not for adolescents. Conversely, no significant correlation was observed for either of the separate groups in the mid-cingulate cortex.

3.8. Correlation between neural and behavioral sensitivity to peer confidence

We wanted to further validate the link between neural activity and individual differences in computational parameters. We limited our brain–behavior correlation analysis to peer confidence, as age differences in sensitivity to own certainty did not persist at the time of information integration and were opposite in direction across neural and behavioral levels. To this aim, we extracted mean subject-level BOLD contrast estimates from 3 mm-radius spherical ROIs centered on the peak voxel of each of the three group-level clusters showing age-related differences in sensitivity to peer confidence. We identified these clusters using the model-based GLM as described above. However, as this GLM already includes the subject-specific parametric values from individual model fits, the resulting beta estimates are incomparable across participants. We therefore extracted BOLD contrast estimates from the model-free GLM in which peer confidence was modeled as a parametric modulator defined identically across participants (i.e. condition-based). We then correlated the BOLD contrast estimates with each individual’s computationally derived θ_slope, which reflects behavioral sensitivity to peer confidence. As expected, this revealed age-group difference in brain-behavior correlations (ventromedial PFC: Fisher’s z = 2.72, p = 0.007; anterior medial PFC: Fisher’s z = 1.92, p = 0.055; midcingulate cortex: Fisher’s z = 2.18, p = 0.029; Fig. 7), with adolescents showing no significant correlation between BOLD signal and θ_slope (ventromedial PFC: r = -0.24, p = 0.102; anterior medial PFC: r = -0.16, p = 0.277; midcingulate cortex: r = -0.19, p = 0.211), but adults showing a (marginally) significant positive correlation (ventromedial PFC: r = 0.29, p = 0.025; anterior medial PFC: r = 0.22, p = 0.096; midcingulate cortex: r = 0.25, p = 0.060). These findings suggest that the link between neural and computational sensitivity to peer confidence strengthens with age.

4. Discussion

In this study, we examined the neurocognitive processes of social information use, and compared whether these differed between adolescents and adults. Overall, we found that participants across age groups relied more on their own estimates than on social information, but flexibly adjusted their behavior depending on their own certainty and the confidence expressed by a peer. Adolescents and adults showed broadly similar decision-making processes, though a combination of model-free and model-based analyses revealed that adolescents were less sensitive to their own certainty and peer confidence. The neural data showed overlapping brain responses across age groups in regions previously linked to uncertainty and social cognition, including the dlPFC, mPFC, insula, caudate nucleus and TPJ. However, adolescents showed heightened neural responses to their own certainty in anterior cingulate and medial prefrontal regions, though this difference was no longer present at the stage of information integration. In contrast, adults exhibited stronger neural responses to peer confidence in the vmPFC. Below, we first discuss the behavioral patterns of social information use, before turning to the neural data.

Roughly similar to previous studies (Gradassi et al., 2022a, Molleman et al., 2019, Yaniv, 2004), we find that people rely more on their personal information than on the social information, a phenomenon known as egocentric discounting. Independent of age, our data revealed that participants used more social information when they were uncertain themselves and when the peer reported to be more confident. Moreover, own certainty and peer confidence interacted, such that social information use was more strongly impacted by peer confidence when people were uncertain, to a degree where social information use well exceeded 0.5. This interactive effect is in line with recent findings (Gradassi et al., 2022a), and shows that the egocentric bias can be overcome in situations where social information is most likely to be beneficial. Our model-based results were in line with our model-free results. For both groups, behavior was best explained by a model that combined Bayesian information processing based on precision weighing of the personal and social information with a bias to stay with the first estimate. This means that participants assigned a higher weight to their personal information in the certain condition, and they assigned a higher weight to the social information when the peer reported to be more confident. A stay bias that was independent of this Bayesian process allowed for participants’ tendency to stick with their initial estimate, in line with earlier work finding that social information use can best be described by a combination of Bayesian decision-making and simple heuristics (Molleman et al., 2020). This suggests that social information, especially low quality information, is often ignored, even though it could in principle still improve participants’ estimates when sufficiently taking into account the level of precision.

Adolescents and adults did not differ in terms of overall social information use. This is in contrast with the intuitive and empirically-informed prediction that adolescents would use more information. Previous experimental studies did find that adolescents were more influenced by social information than adults (Andrews et al., 2021, Somerville, 2013). Interestingly, when approximately half an hour after the mushroom task the same individuals performed a similar perceptual decision-making task (Berlin Estimate Adjustment Task; (Molleman et al., 2019)), albeit it without any certainty or confidence manipulations, we found that adolescents used more social information than adults. This suggests that adolescents are more inclined to use social information when reliability cues are absent. In contrast, adults may be more hesitant to rely on social information unless there is a clear indication of its reliability or a strong need for additional information. Alternatively, we speculate that the complexity of the current task (e.g. longer trial duration, multiple screens to display personal and social information, multiple levels of own certainty and peer confidence) could have led the younger group of participants to rely more on one particular set of information, conceivably their personal information, rather than integrating all information available to them to the extent they would have done during a less demanding task. Thus, while our findings did not replicate a group-level difference in overall social information use, they do not necessarily contradict the notion of adolescence as a period of heightened social sensitivity. Instead, our data suggest that adolescents' social information use might be more context-dependent: in the absence of explicit reliability cues, they may be more inclined than adults to rely on social information, or conversely, under increased task complexity, they may default more strongly than adults to their own information.

Compared to adults, adolescents showed a smaller effect of the own certainty condition on social information, mainly driven by adolescents’ lower social information use in the uncertain condition compared to adults. The effect of peer confidence did not differ across age groups. Both adolescents and adults used more social information when the peer reported to be more confident, similar to people’s response to indicators of intelligence and social status (Gradassi et al., 2022b, Helms et al., 2014, da Silva Pinho et al., 2021). Our computational model provided a deeper understanding of the cognitive process by breaking it down into distinct components, offering more nuanced insights compared to the model-free analysis. While adolescents exhibited a smaller difference between the certain and uncertain conditions, consistent with the model-free findings, the model revealed that this difference was driven by adolescents behaving as if they were less certain than adults in the certain condition. Additionally, the model highlighted that adolescents were less sensitive to peer confidence levels while showing no differences in stay bias compared to adults. The apparent contrast between the model-free and model-based findings can be understood by examining how the Bayesian model integrates subjective uncertainty. According to the model-based results, in the uncertain condition, adolescents and adults did not differ in how many mushrooms they subjectively perceived in their own estimate. However, adolescents perceived the peer's estimate as less precise than adults did, especially when the peer expressed high confidence (Fig. 8, top). In Bayesian terms, this means that adolescents assigned a wider (i.e., less precise) distribution to the peer information. As a result, they gave less weight to the peer's input when updating their estimate, which matches the lower social information use observed in the model-free data. In contrast, in the certain condition, adolescents subjectively perceived fewer mushrooms than adults did in their own estimate. This implies they assigned less precision (i.e., a wider prior) to their own information. At the same time, they continued to perceive the peer's estimate as less precise than adults did (Fig. 8, bottom). This led to a situation where both sources, personal and social, were seen as relatively imprecise by adolescents, resulting in an update that was similar in size to that of adults. This is consistent with the lack of a group difference in social information use in the certain condition as seen in the model-free results. The key difference lies in how certain participants were about their final (second) estimate: in the certain condition, adolescents showed greater uncertainty in their second estimate than adults, as indicated by the wider posterior distribution (E2). This suggests that while the amount of adjustment may appear similar in model-free terms, adolescents were still internally less confident in their final judgments, a difference that is only captured by the model-based analysis. Together, these findings highlight how computational modeling can reveal latent cognitive differences that remain hidden in behavioral outcomes.

Consistent with our results, previous studies have indeed shown that adolescents often display behavior indicative of greater uncertainty compared to adults in learning tasks (Javadi et al., 2014, Jepma et al., 2020). These observations may be related to still maturing metacognitive skills of adolescents (Moses-Payne et al., 2021, Weil et al., 2013). While our data suggest that adolescents and adults do not differ in how they translate their own performance into subjective confidence reports, adolescents may nonetheless be less accurate in appraising the uncertainty in the environment. This may account for their reduced ability to perform precision-weighting and to accurately differentiate between varying levels of certainty, as well as peer confidence statements (Reiter et al., 2021). An alternative explanation is that adolescents are less ambiguity averse than adults (Tymula et al., 2012, Van Den Bos and Hertwig, 2017), and may therefore feel less need to resolve ambiguity by incorporating peer information when uncertain about their own estimate. This could result in lower social information use specifically in the uncertain condition. However, this explanation appears to be less consistent with the observed computational results: if adolescents were simply less ambiguity averse, we would expect them to assign relatively high precision to their own estimates and rely less on the peer. Instead, the computational model shows that adolescents behaved as if they had lower subjective certainty about both their own and the peer’s information, particularly in the certain condition. This pattern suggests not a reduced need to resolve ambiguity, but rather a difference in how adolescents weigh uncertainty.

Together, our behavioral findings suggest that while adolescents and adults both rely on social information influenced by their own certainty and peer confidence, adolescents show a smaller distinction between different levels of own certainty and peer confidence, possibly due to ongoing neural and metacognitive development. Despite these differences, the core processes of social information use appear similar across age groups, combining Bayesian decision-making with heuristic biases.

To gain more insights into the neural mechanisms underlying sensitivity to own certainty and peer confidence, we compared the neural activity of adolescents adults using fMRI. Firstly, this revealed a large overlap in neural processing between the two groups. We found that neural activity correlated positively with own certainty mainly in the dorsal parts of the PFC, the anterior insula and the caudate nucleus. A negative correlation, on the other hand, was mainly present in the more anterior and ventral parts of the mPFC, the posterior parts of the insula and the Rolandic operculum and areas within the TPJ. Many of these areas have indeed been implicated in the cognitive processing of uncertainty-related information. For example, the anterior insula and dlPFC have been linked to processing risk and uncertainty (Badre et al., 2012, Feldmanhall et al., 2019, Singer et al., 2009), while vmPFC activation is thought to reflect confidence signals (Gherman and Philiastides, 2018, Lebreton et al., 2015, Rouault et al., 2023), though these signals can sometimes manifest in opposite directions. In addition to processing confidence signals, activation in the vmPFC in response to uncertainty, along with activation in the TPJ, may suggest that participants are proactively preparing to process the upcoming peer information when their own evidence base is limited, aligning with the TPJ’s role in social processing and attention to peers (Jin et al., 2023, Sul et al., 2017, Van Overwalle, 2011). Involvement of the caudate nucleus in processing both own certainty and peer confidence might be due to its role in value signaling (Delgado et al., 2004; Grahn et al., 2008), with more information in the certain condition and peer information associated with higher confidence levels being more valuable. Importantly, a similar set of regions whose activity correlated with own certainty and peer confidence as set by the task manipulations also correlated with the model-based estimates of own certainty and peer confidence. This suggests that the neural patterns align with our Bayesian model of social information use, even though our model may be too simplistic to capture the entire decision-making process.

When comparing age differences in neural activity, we observed larger neural responses to differences in own certainty in adolescents compared with adults, including in the Rolandic operculum, the anterior regions of the mPFC and cingulate cortex and the supplementary motor area. This is surprising, as we found stronger sensitivity at the behavioral level for adults. However, the increased neural responsivity in adolescents was no longer present at the time of the peer information, suggesting that this strong initial response in adolescents was not utilized when integrating personal and social information. This might point at poorer metacognition in adolescence compared to adulthood (Moses-Payne et al., 2021, Weil et al., 2013), leading to an aberrant translation of initially available neurocognitive information to decisional output in adolescence, which might in turn stem from, subsequently affecting the search or use of additional information (Schulz et al., 2023).

Although we were initially quite agnostic about the cognitive differences in sensitivity to peer confidence between adults and adolescents, we found that adults showed a stronger response than adolescents, particularly in the vmPFC, which aligns with our model-based results. As mentioned earlier, this region is heavily implicated in both confidence processing and social cognition. Lower responsiveness in the vmPFC might therefore mean that adolescents possess a lower ability to distinguish between varying levels of peer confidence. This could result from decreased recruitment of Theory-of-Mind processes to modulate attention to peer information and aligns with prior research showing that learning for others but not for the self, as reflected by prediction error coding in the vmPFC, showed an age related increase during adolescence (Westhoff et al., 2021). However, other regions related to social cognition and especially Theory-of-Mind, including the TPJ, do not exhibit differential neural activity between adolescents and adults. Another possible explanation comes from findings that the vmPFC plays a crucial role in the integration of information (Campbell-Meiklejohn et al., 2017, Kahnt et al., 2011, Park et al., 2020). Notably, brain–behavior correlations showed that BOLD signal in the vmPFC and anterior mPFC correlated with individual differences in model-based sensitivity to peer confidence only in adults and not in adolescents. Thus, our findings might indicate that this region is functionally less developed or active in adolescents compared to adults when it comes to integrating personal and social information.

The strength of the current design, combined with our computational model, lies in its ability to disentangle the various components of social information use under uncertainty. However, it is not without its limitations. One limitation is the order of the presentation of personal and peer information. It is hard to distinguish the distinct neurocognitive contributions of one's own certainty, peer confidence, and the integration of the two at the second time point, as participants may begin integrating the information as soon as they see the peer's input. Future studies could vary whether participants receive their personal or the peer's information first, enabling researchers to assess how each type of information is processed independently before integration. Additionally, such a design could reveal whether the sequence of information presentation affects the overall decision-making process. Another limitation of our study is how information was presented across conditions. In the uncertain condition, participants were given a small amount of clear and easily memorable information, while in the certain condition, they received a larger volume of more reliable information, presented rapidly. This introduces a potential confound between inferential certainty (where the uncertain condition was less certain) and perceptual certainty (where participants might feel more confident about what they saw in the uncertain condition, despite it being harder to interpret). This distinction could have influenced the neural signals of certainty or uncertainty, potentially explaining why some of our findings contrast with the directions reported in other neuroimaging studies. Future research should consider adjusting how information is presented, such that inferential certainty is manipulated while perceptual certainty is aligned across conditions.

An interesting avenue for future research is to more accurately pinpoint belief precision during Bayesian social information use. Even though our results suggest that adolescents exhibit lower sensitivity to certainty and confidence due to a more limited integration of precision signals, future studies could test this in more detail at the neural level by employing a neuroimaging design in which the timings are optimized to model neural activity in response to each successive piece of personal or social information (Skowron et al., 2024). This could be combined with a larger and more varied range of ages within the adolescent sample to assess any continuous effects of age to evaluate how and when neurocognitive mechanisms of social information use and precision weighing develop during adolescence. Ideally, a longitudinal design would be employed to preclude any cohort effects, such that participants are tested multiple times from early adolescence through late adolescence or early adulthood.

While we did not directly assess how participants interpreted the peer confidence ratings, the peer had sampled from different fields with potentially different numbers of mushrooms, which suggests that peer confidence reflected informational resources. However, we acknowledge that we cannot be sure whether participants interpreted higher peer confidence as better informational resources, general expertise, or even trait characteristics. This makes it difficult to determine whether social information use in our task was driven primarily by informational influence (aiming to improve one’s own accuracy) or by normative influence (aiming to align with a confident peer). Importantly, these motives may differ across development. While adults may be more selective in whose information they trust based on perceived expertise (Hofmans and van den Bos, 2022), adolescents may be more sensitive to normative social influence and place greater weight on social information when it comes from peers who are similar to them in age or identity, or with whom they share a close relationship. A related question is how individuals gradually learn about the confidence or trustworthiness of a peer, rather than being provided with an explicit rating as in the current study. This process of learning and integrating social information is likely to be influenced by various factors, including an individual's prior experiences (e.g. interactions with trustworthy or untrustworthy peers) (Olsen et al., 2019). Future studies could explore how prior experiences shape the interpretation of peer information.

5. Conclusion

In sum, our findings contribute to the growing understanding of how adolescents and adults process and integrate social information. While both age groups exhibit similar core mechanisms, such as Bayesian decision-making combined with heuristic biases, notable differences emerge in how own certainty and peer confidence influence their behavior and neural responses. Adolescents appear to show reduced sensitivity to variations in their own certainty and peer confidence, possibly due to ongoing neural and metacognitive development, which may hinder their ability to effectively integrate precision signals with the relevant information.

Introduction

Adolescents are often more focused on their peers and more influenced by them than at other times in life. This peer influence, or using social information from others, can be helpful. It allows adolescents to learn how to act well and ethically as they navigate new social situations. Copying what others do can also speed up learning, preventing the need for trial-and-error, which can be time-consuming or risky.

However, not all information from peers is equally valuable, and there are times when one should not be swayed by others' opinions. A key skill for adolescents to develop is knowing when to rely on peer advice. The usefulness of social information depends on two main factors: a person's current knowledge and the expertise of the source. When a person is in a new or uncertain situation with little knowledge of what to do, there is a greater tendency to look to others for guidance. This might explain why adolescents, who are constantly encountering new and changing social environments, are generally more open to peer information.

Research has shown that adolescents tend to seek out and use social information more than adults. While both adults and adolescents use more social information when they are feeling uncertain, it is not yet clear how adolescents' sensitivity to uncertainty compares to adults.

Beyond a person's own level of uncertainty, the perceived expertise of others also plays an important role in deciding how much weight to give their information. In everyday life, objective measures of expertise are often unavailable. Instead, people frequently communicate how confident they are. Both children and adults are more likely to use social information from peers who express high confidence. Although adolescents are generally more focused on their peers, and they do use social information based on a peer's intelligence or social standing, their response to a peer's confidence level has not been fully explored.

Previous studies suggest that people use social information in a way that is similar to Bayesian updating. This means they weigh different pieces of information based on how precise or reliable they seem. For example, if a person is very sure about their own experiences, they will give more importance to that personal information than to social information. On the other hand, if a peer states they are very confident, this is often interpreted as highly precise information, leading to a stronger change in belief based on what the peer says. Both adolescents and adults also use simpler decision rules, such as copying peers if many of them do the same thing, or if peers act consistently. The development of these complex reasoning processes and simpler rules throughout adolescence remains unclear.

These changes in how decisions are made during adolescence may be linked to ongoing brain development. In adults, uncertainty and precision are represented across a wide network of brain regions, including areas involved in higher-level thinking like the medial prefrontal cortex (mPFC) and the anterior cingulate cortex (ACC). Other brain areas, such as the temporoparietal junction (TPJ) and superior temporal sulcus (STS), are more specifically involved in understanding others' thoughts and intentions. The integration of personal and social information has been observed in regions like the ventromedial prefrontal cortex (vmPFC).

During adolescence, the brain's prefrontal cortex (PFC) continues to mature, which could lead to less accurate estimations of precision and poorer awareness of one's own thought processes. This might mean adolescents are generally more uncertain and less able to respond to varying levels of their own uncertainty or a peer's confidence, making them more easily influenced by peers or more likely to use simple decision rules. However, the strong focus on peers during adolescence, possibly due to increased activity in reward-related brain regions, could also make them more sensitive to peer cues. This study aims to compare how adolescents and adults use social information when faced with uncertainty, and to examine the brain activity linked to these processes.

Methods

A study was conducted to understand how adults and adolescents use social information when uncertain. One hundred four adults (aged 22–45) and 105 adolescents (aged 12–18) were recruited. Participants were right-handed, spoke Dutch, had normal vision, and had a university or pre-university level education. The study involved an online questionnaire and an in-person session with an MRI scanner. The MRI session included a main task focused on using social information, as well as other cognitive tasks. After accounting for incomplete data, age range, and behavioral or MRI quality issues, 68 adults and 64 adolescents were included in the behavioral analysis, and a subset of these (58 adults, 47 adolescents) were included in the brain activity analysis. All participants gave written consent, with parents providing consent for those under 16.

The main task was an estimation game where participants acted as community members estimating the percentage of blue mushrooms in a forest. They first made an initial estimate after seeing a varying number of mushrooms in surrounding fields. The number of mushrooms seen determined their "own certainty" (more mushrooms meant more certainty). After making their initial estimate, participants were shown information from a "peer" (another participant from a previous study). This included the peer's confidence level (low, medium, or high) and their estimate. Participants then had the opportunity to revise their initial estimate. The task was designed so that the peer's estimate would always be somewhat different from the participant's first estimate, allowing for adjustments. No feedback was given on their estimates, but participants could earn a bonus based on accuracy.

The task involved 75 trials, divided into three runs, lasting about 30 minutes. Trials were structured to vary own certainty (certain, uncertain) and peer confidence (low, medium, high). Before the main task, participants received instructions and practiced, including rating their own confidence, to ensure they understood the game. This practice also helped identify participants who might not have grasped the core concept, allowing for further explanation.

Brain imaging data was collected using a 3 Tesla MRI scanner. High-resolution structural images of the brain were taken for anatomical reference, and functional images measured blood-oxygen level-dependent (BOLD) contrast to detect brain activity. The images underwent standard preprocessing steps to correct for various distortions and align them for analysis. Participants with excessive head motion or other quality issues were excluded from the fMRI analysis.

The main measure of behavior was "social information use," which was the extent to which participants adjusted their first estimate towards the peer's estimate. A value of 0 meant no adjustment, and 1 meant fully adopting the peer's estimate. Trials where participants didn't understand the task, responded randomly, or showed minimal variation in social information use were excluded. Linear mixed effects regression models were used to analyze how own certainty, peer confidence, and age group affected social information use. Control analyses checked if performance differences or how confidence aligned with performance could explain age group effects.

A computational modeling approach was used to understand the underlying thought processes. Several models, increasing in complexity, were developed to explain how participants combined their own information with peer information. These models assumed a process similar to Bayesian updating, where information is weighted based on its perceived reliability. The best-fitting model included parameters to adjust how much weight was given to one's own information based on certainty, and how much weight was given to the peer's estimate based on their confidence. It also included a "stay bias" parameter, which accounted for participants sometimes choosing to stick with their initial estimate regardless of peer input. Model recovery and parameter recovery procedures were performed to ensure the models and their estimated parameters were reliable.

Finally, brain activity data from the fMRI scanner was analyzed. This involved modeling BOLD responses to different events in the task. Parametric regressors were used to link brain activity to "model-based own certainty" (how certain participants were based on the computational model) and "model-based peer confidence" (how influential peer confidence was according to the model). Separate analyses looked at age group differences in brain activity. The focus was on brain regions previously linked to uncertainty and social cognition, such as parts of the prefrontal cortex, insula, temporoparietal junction, and striatum. Brain-behavior correlations were also conducted to link individual differences in computational parameters with brain activity.

Results

Participants in both age groups made initial estimates that accurately reflected the actual percentages of mushrooms, indicating a clear understanding of the task. On average, participants relied more on their own estimates than on social information, showing an "egocentric bias." Across all participants, social information use was higher when individuals were uncertain about their own estimate and when the peer expressed higher confidence. There was also an interaction: the effect of peer confidence was stronger when participants were uncertain themselves. This suggests that people will adjust their egocentric bias when social information is most likely to be helpful.

When comparing age groups, there was no overall difference in how much adolescents and adults used social information. This finding contrasts with some earlier studies that showed adolescents using more social information. However, when these same individuals performed a similar task without certainty manipulations later, adolescents did use more social information than adults. This suggests that the presence of explicit certainty and confidence cues in this specific task might have influenced how adolescents used social information, possibly leading them to rely more on their own information in this complex scenario.

A significant age difference was found in how sensitive participants were to their own certainty: adolescents were less affected by their own certainty than adults, particularly in the uncertain condition, where they used less social information than adults. However, the effect of peer confidence on social information use did not differ between age groups. Exploratory analysis of self-reported confidence during practice showed that adolescents felt more confident than adults when they were in the uncertain condition, but not in the certain condition. No age-related differences were found in task performance or in how well confidence correlated with performance. This suggests that differences in how adolescents use social information based on their own certainty are not simply due to differences in performance or in judging their own accuracy.

Computational modeling helped to further clarify these findings. The best-fitting model for both age groups combined Bayesian information processing (weighing personal and social information based on perceived precision) with a "stay bias" (a tendency to stick with the initial estimate). The model showed that participants amplified their subjective certainty more in the uncertain condition compared to the certain condition. Furthermore, higher peer confidence led participants to act as if the peer had seen more mushrooms, thus giving more weight to the peer's estimate. The stay bias was influenced by peer confidence, with participants less likely to stick with their first estimate when peer confidence was high.

When comparing the computational parameters between age groups, adolescents showed a significantly lower parameter for their own certainty in the certain condition compared to adults. This means adolescents behaved as if they were less certain than adults when they had more personal information. As a result, the difference in model-based certainty between the certain and uncertain conditions was smaller for adolescents. Adolescents also showed less sensitivity to peer confidence: for each increase in peer confidence, adolescents acted as if the peer had seen fewer additional mushrooms compared to adults. There was no age difference in the stay bias. This suggests that while adolescents perceive similar levels of certainty as adults in uncertain situations, they discount the reliability of both their own information (when highly certain) and peer information (especially at higher confidence levels) more than adults. This leads to a smaller overall shift in their estimates and a relatively lower confidence in their final decisions.

Brain imaging results showed common patterns of activity across both age groups. When making their first estimate, brain activity positively correlated with own certainty (meaning more activity when more certain) in areas like the dorsomedial and dorsolateral prefrontal cortex, anterior insula, and left caudate nucleus. A negative correlation (more activity when less certain) was found in areas like the supplementary motor area, medial prefrontal cortex, posterior insula, and bilateral temporoparietal junction. These regions are known to be involved in processing uncertainty and social cognition. Neural activity related to peer confidence at the time social information was presented showed positive correlations in the left precentral gyrus and caudate nucleus. These findings align with the brain activity expected from the computational model.

Comparing neural activity between adolescents and adults revealed age-related differences. Adolescents showed stronger negative correlations with own certainty in the anterior cingulate cortex, medial prefrontal cortex, and supplementary motor area when making their first estimate. This suggests adolescents were more neurally sensitive to their own certainty at this initial stage. However, this neural difference was not present at the later stage when personal and social information were integrated. In contrast, adults showed stronger neural responses to peer confidence than adolescents, particularly in the anterior and ventral parts of the medial prefrontal cortex (vmPFC). For adults, there was a positive correlation between vmPFC activity and peer confidence, which was not observed in adolescents. This suggests that adults may use the vmPFC more actively to process and integrate peer confidence signals. Furthermore, the link between brain activity in vmPFC regions and an individual's behavioral sensitivity to peer confidence (as measured by the computational model) was stronger in adults than in adolescents, indicating a developing brain-behavior relationship with age.

Discussion

This study investigated how adolescents and adults process and use social information. It found that while both groups rely more on their own estimates, they adjust this behavior based on their personal certainty and a peer's confidence. This adjustment is flexible, especially when social information is likely to be helpful. The findings suggest that decision-making involves a combination of Bayesian processing, which weighs information based on its reliability, and simpler decision rules, such as a bias to stick with one's initial estimate.

Contrary to some prior research, this study did not find that adolescents used more social information overall than adults. This might be because the task was more complex and included explicit cues about certainty and confidence, which could have led adolescents to rely more on their own information. The study highlights that adolescents' use of social information might depend on the specific context, such as the absence of reliability cues or the complexity of the task.

A key finding was that adolescents were less sensitive to changes in their own certainty compared to adults. Computational modeling revealed that adolescents behaved as if they were less certain of their own information when they should have been highly certain, and they also gave less weight to peer estimates, particularly when peers expressed high confidence. This suggests adolescents might have a less precise way of evaluating certainty and confidence signals, leading to smaller distinctions between different levels of reliability. These differences in how adolescents weigh uncertainty could be linked to their still-developing metacognitive skills, which involve awareness and control over one's own thought processes.

Brain imaging showed shared brain networks involved in processing certainty and social information across both age groups. Areas like the prefrontal cortex, insula, and caudate nucleus were active, which are known to be involved in assessing risk, confidence, and value. Importantly, the patterns of brain activity aligned with the computational model, supporting the idea that the brain uses these Bayesian-like processes.

However, some differences in brain activity were observed between age groups. Adolescents showed a stronger initial brain response to their own certainty in regions like the anterior cingulate cortex and medial prefrontal cortex. This stronger response, however, did not carry over to the later stage of integrating information, suggesting that adolescents might not fully use this initial neural signal in their decision-making. In contrast, adults showed a stronger brain response to peer confidence in the ventromedial prefrontal cortex (vmPFC), a region important for confidence and social cognition. This suggests that the vmPFC might be more functionally developed or active in adults for integrating social information, and that the brain-behavior link between vmPFC activity and sensitivity to peer confidence strengthens with age.

While this study offers valuable insights, it has limitations. The order of presenting personal and peer information makes it challenging to separate brain activity related to each, and how the task manipulated certainty could be refined in future studies. Future research could also focus on more detailed brain imaging, a wider and more varied age range of adolescents, and longitudinal designs to better understand the continuous development of social information use and certainty processing. Additionally, exploring how individuals learn about peer trustworthiness over time, rather than being given explicit confidence ratings, would be beneficial.

Conclusion

In conclusion, this study reveals that while adolescents and adults generally follow similar decision-making patterns when using social information, there are important developmental differences. Adolescents show reduced sensitivity to variations in their own certainty and peer confidence, which may be linked to ongoing brain development and how they integrate information. These findings improve our understanding of how social influence works across different stages of adulthood.

1. Introduction

Adolescents are generally more focused on peers and more susceptible to peer influence than adults. This influence, also known as using social information, can be helpful. For example, adolescents can learn how to behave appropriately and morally in new situations by observing their peers. Copying others' behavior can also make learning faster, avoiding the need for trial-and-error.

However, not all peer information is equally valuable, and it is important to know when to rely on others. The usefulness of social information depends on a person's current knowledge and the source's expertise. When individuals are in new or uncertain situations, they tend to rely more on what others do. Adolescence is a period filled with new and changing social situations, which may explain why adolescents often seek more information and use more social information than adults. People also tend to use social information from peers who seem more confident. While adolescents are generally focused on their peers and consider factors like a peer's intelligence or social standing, their response to a peer's confidence level has not been fully explored.

Research suggests that using social information often involves a process similar to Bayesian updating. This means people give more weight to information they perceive as more certain or reliable. If someone is very confident in their own experience, they will prioritize that over social information. Conversely, if a peer seems very confident, their input will be given more consideration. In addition to this logical weighing of information, both adolescents and adults also use simpler decision-making rules. These might include copying peers if they represent a majority or act consistently. It is not yet fully understood how these logical, Bayesian processes and simpler rules develop during adolescence. While complex reasoning skills improve with age, some decision biases may change in different ways.

These shifts in decision-making may be linked to ongoing changes in the brain during adolescence. Many brain areas are involved in processing uncertainty and confidence, including parts of the prefrontal cortex, insula, and striatum. Other areas, like the temporoparietal junction (TPJ), are involved in understanding others' thoughts and beliefs (Theory of Mind). As the prefrontal cortex continues to mature during adolescence, it might lead to less accurate estimates of certainty. This could make adolescents more generally uncertain or less sensitive to different levels of their own certainty and peer confidence, potentially making them more susceptible to peer influence or to using simple decision rules. The development of social understanding (Theory of Mind) and brain regions like the TPJ could also affect how adolescents respond to peer cues.

This study aims to compare how adolescents and adults use social information when faced with uncertainty. A task was used where participants could adjust their initial estimate based on a peer's response, with varying levels of personal certainty and peer confidence. Researchers predicted that adolescents would generally use more social information and be less sensitive to their own certainty. The study also explored brain activity using fMRI to see how responses to personal and peer-related certainty differ between age groups, focusing on areas like the prefrontal cortex, insula, TPJ, and striatum.

Introduction

Adolescents often focus more on their peers and are more open to peer influence. This influence can be very helpful, as peers can provide information on how to act appropriately and morally in changing situations. Following others can also speed up learning, preventing the need to learn everything through trial and error.

However, not all information from peers is equally valuable, and people should not always be swayed by their opinions. Learning when to rely on peers is an important goal for adolescents. The value of using social information depends on a person's current knowledge and the peer's expertise. When people are in new or uncertain environments, they tend to rely more on others' behavior. Adolescence is a time of many new and changing social situations, which may explain why teenagers are often more open to peer information. Still, it is not yet fully clear how adolescents' sensitivity to uncertainty compares to adults.

Beyond personal uncertainty, a peer's confidence level also plays an important role. People, both young and old, tend to use social information more when a peer seems very confident. While adolescents generally focus more on peers and use social information based on a peer's intelligence or social standing, their reaction to a peer's confidence has not been fully explored.

Decision-making often involves balancing personal information with social information. This process is similar to how people update their beliefs by giving more importance to information they are more certain about. So, if a person is very sure about their own experience, they will trust that more than a peer's opinion. Conversely, if a peer seems very confident, that opinion might be given more weight. People also use simpler decision rules, such as copying a peer if many peers agree or if a peer acts consistently. How these complex and simple decision strategies develop during adolescence is still being investigated.

Changes happening in the brain during adolescence may affect how these decisions are made. Brain areas linked to certainty, confidence, and social interaction are still maturing. This could mean adolescents are less accurate at judging how certain they are, making them more likely to be influenced by peers or to use simple decision rules. Alternatively, their strong focus on peers might make them more responsive to peer cues, driven by certain brain regions.

This study aims to compare how adolescents and adults use social information when they are uncertain. The researchers predict that adolescents will use social information more often overall and be less sensitive to their own certainty. The study also investigates how brain responses to personal and peer certainty differ between these age groups.

Methods

Participants

The study included 104 adults (ages 22–45) and 105 adolescents (ages 12–18). Participants were recruited from a university research system, social media, and local schools. All participants were right-handed, Dutch-speaking, had normal vision, and met specific educational requirements for their age group. They were screened to ensure they could safely undergo an MRI scan. Out of these, 68 adults and 64 adolescents were included in the behavioral analysis after some participants were excluded due to incomplete data, age outside the defined range, or not understanding the task. A smaller subset of these participants (58 adults and 47 adolescents) were included in the fMRI analysis after additional quality checks for brain scan data. All participants and, for those under 16, their parents, provided written consent.

Social information use task

Participants played an estimation game designed to measure how they used social information based on their own certainty and a peer’s confidence. In the game, participants were told to estimate the percentage of blue mushrooms in a forest each "day." To do this, they first saw mushrooms in five fields. On some days, they saw many mushrooms (45 total, making them more certain), and on other days, they saw few (5 total, making them less certain). Participants gave an initial estimate using a slider.

After their first estimate, they saw information from a "peer" (a previous participant). This included the peer's confidence level (low, medium, or high) followed by the peer's actual estimate. Participants could then adjust their own estimate. Peer estimates and confidence ratings came from real past participants. These peer estimates were chosen to be slightly different from the participant's first estimate, encouraging adjustment. Participants completed 75 trials and received a bonus based on how close their final estimate was to the true percentage. Before the main task, participants completed practice rounds to ensure they understood the game, including rating their own confidence to understand what the peer's confidence ratings meant.

MRI acquisition and preprocessing

Brain scans were performed using a 3 Tesla MRI scanner. High-resolution structural images were taken for anatomical reference, and functional images were acquired to measure blood-oxygen level-dependent (BOLD) contrast, which indicates brain activity. The raw image data was then converted and processed using standard software. This processing included correcting for various distortions, aligning images, and preparing them for statistical analysis to ensure data quality. Participants with excessive head motion or other quality issues in their scans were excluded from the fMRI analysis.

Behavioral data analysis

The main measure was "social information use," which was calculated as how much a participant adjusted their initial estimate towards the peer's estimate. A value of 0 meant no change, and 1 meant completely copying the peer. Trials where participants didn't respond or where social information was unsuitable were excluded. Researchers used a statistical method called linear mixed effects regression to analyze how own certainty, peer confidence, and their interaction affected social information use across all participants. They also included age group in a separate analysis to see if adolescents and adults differed.

Control analyses

To make sure that any differences found between age groups in social information use were not due to differences in how well they performed the task, additional analyses were conducted. Performance was measured by how far off the participant's first estimate was from the true percentage of blue mushrooms. Researchers also checked if adolescents and adults differed in how well their reported confidence matched their actual performance during practice. Another check ensured that differences in how consistently participants responded (residual variance) did not affect the results. Finally, analyses were also run treating age as a continuous number, rather than just two groups, to confirm the findings.

Computational modeling

Mathematical models were used to understand the underlying mental processes guiding participants' decisions. These models aimed to describe how people weigh different pieces of information, such as their own estimate and the peer's estimate, to make a final decision. The models started with a basic framework that assumed participants' estimates followed a certain statistical distribution, updated by observed mushrooms.

Individual modulation of the weight of E1

One set of models explored whether participants adjusted how much weight they gave to their own initial estimate. This involved adding a parameter that could change the perceived total number of mushrooms a participant had seen, effectively changing how certain they felt about their own information. Some models allowed this adjustment to be different when they were in an "uncertain" condition versus a "certain" condition.

Individual modulation of the weight of the peer estimate

Another set of models investigated whether participants adjusted the weight they gave to the peer's estimate. This involved a parameter that could change the assumed number of mushrooms the peer had seen. These models explored if this adjustment was fixed or if it changed based on the peer's reported confidence level, either in a linear way (steadily increasing with confidence) or a non-linear way (allowing for different impacts at low, medium, or high confidence).

Stay bias

Researchers observed that participants sometimes just stuck with their first estimate without changing it at all. To account for this, a "stay bias" parameter was added to the models. This parameter represented a tendency to not change one's mind, and it could be non-selective (the same across all trials) or could vary based on the participant's own certainty or the peer's confidence. The best model allowed this stay bias to decrease exponentially as peer confidence increased, meaning people were more likely to change their mind when the peer was more confident.

Model fitting procedure

Each model's predictions were compared against each participant's actual decisions. The goal was to find the parameter values for each model that best matched the observed behavior. To compare the different models, researchers used the Bayesian Information Criterion (BIC), which favors models that fit the data well without being too complex. The model with the lowest BIC was considered the best fit. After identifying the best model, the estimated parameters for adolescents and adults were compared statistically to find differences in their cognitive processes.

Model recovery and parameter recovery

To ensure the models were reliable, a "model recovery" procedure was performed. This involved simulating data based on each model and then checking if the correct model was chosen as the best fit. This helped confirm that the chosen model wasn't just fitting random noise. Additionally, "parameter recovery" was done to confirm that the estimated parameter values accurately reflected the true underlying values that generated the simulated data.

fMRI data analysis

Brain activity data from the fMRI scans were analyzed to see which brain regions were active during the task. Researchers used statistical models to link brain responses to different events in the game. Specifically, they looked at brain activity when participants made their first estimate, correlating it with their internal "model-based own certainty." They also examined brain activity when peer information was presented, correlating it with "model-based peer confidence." They created "contrasts" to highlight areas that showed more activity for certainty versus uncertainty, or for higher versus lower peer confidence. Control analyses using simpler, "model-free" conditions (e.g., just the "certain" vs. "uncertain" trial types) were also performed to validate the results. Finally, direct comparisons were made between adolescents and adults to find brain regions with significant differences in activity.

Results

Effect of own certainty and peer confidence on social information use

Participants generally used more of their own information than the peer's information, a common pattern called egocentric bias. On average, they adjusted their initial estimate by about 37% towards the peer's. However, participants across all age groups adjusted their estimates more when they were uncertain about their own information and when the peer reported high confidence. There was also an interaction: the peer's confidence had a stronger impact when participants were already uncertain, sometimes even leading them to rely more on the peer than on their own initial idea.

Age-related differences in sensitivity to own certainty and peer confidence

The study found no significant overall difference in how much social information adolescents and adults used. However, adolescents were less affected by their own level of certainty than adults were. Specifically, adolescents used less social information than adults when they were in an uncertain situation, but there was no difference when they were in a certain situation. The impact of peer confidence on social information use did not differ between age groups. Interestingly, in a separate, simpler task given to the same participants later, adolescents did rely more on social information than adults. This suggests that the presence of certainty and confidence cues in this more complex task might have changed adolescents' typical tendency to use more social information. Furthermore, adolescents reported feeling more confident in uncertain situations compared to adults. Despite these behavioral differences, performance on the task (how accurate their first estimate was) did not differ between age groups, nor did their ability to align their confidence with their performance.

Computational mechanisms of social information use

Mathematical modeling helped to understand the thinking processes involved. For both age groups, the best model suggested that participants weighed their personal information, adjusted how much they considered the peer's estimate based on confidence, and also had a "stay bias" (a tendency to stick with their initial answer). This stay bias became less common as the peer's confidence increased, meaning participants were more likely to change their minds when a peer was highly confident. This model combines a careful weighing of information with simpler decision rules.

Model recovery and parameter recovery

The checks performed on the models showed that they were reliable. When data was simulated based on a specific model, that model was correctly identified as the best fit. The estimated values for the model parameters also accurately reflected the true values used to generate the simulated data. This confirms that the computational models and the measurements derived from them were sound.

Age differences in computational processes

When comparing the parameters of the best model between age groups, researchers found interesting differences. Adolescents and adults had a similar subjective certainty when they were in the "uncertain" condition. However, adolescents showed lower subjective certainty in the "certain" condition compared to adults. This meant adolescents behaved as if they were less certain about their own information when they had more of it. Additionally, adolescents perceived the peer's estimate as less precise than adults did, especially when the peer expressed medium or high confidence. This suggests that adolescents gave less weight to peer estimates as peer confidence increased, unlike adults who gave more weight. There was no significant difference in the "stay bias" parameter, indicating both groups were equally likely to consider changing their minds. These computational results offer a deeper look at how adolescents weigh information, suggesting they might be generally less precise in their judgments of both their own and others' certainty.

Neural activity in response to model-based own certainty and peer confidence across age groups

Brain scans revealed common patterns of activity across both age groups. Areas like the dorsomedial and dorsolateral prefrontal cortex (PFC), anterior insula, and left caudate nucleus showed increased activity when participants were more certain about their own information. Conversely, areas like the medial PFC, posterior insula, and parts of the temporoparietal junction (TPJ) showed more activity when participants were uncertain. When social information was presented, activity in the left precentral gyrus and caudate nucleus increased with peer confidence. These brain regions are known to be involved in processing uncertainty, confidence, and social cues.

Age differences in neural activity in response to model-based own certainty and peer confidence

When comparing brain activity between adolescents and adults, adolescents showed stronger neural responses to their own uncertainty in areas like the anterior cingulate cortex, medial PFC, and supplementary motor area. This suggests adolescents initially reacted more strongly to their own uncertainty at a neural level, which contrasts with their behavioral tendency to be less sensitive to their own certainty. However, this heightened neural response in adolescents was no longer evident when they were integrating personal and social information. In contrast, adults showed stronger brain responses to peer confidence, particularly in the ventromedial PFC (vmPFC), a region important for confidence processing and social cognition.

Correlation between neural and behavioral sensitivity to peer confidence

To further understand the connection between brain activity and behavior, researchers looked at how vmPFC activity related to individual differences in responding to peer confidence. The link between brain activity in the vmPFC and how much people adjusted their behavior based on peer confidence was significantly stronger in adults than in adolescents. In adults, higher vmPFC activity correlated with greater sensitivity to peer confidence, but this relationship was not significant in adolescents. These findings suggest that the functional link between brain activity in key decision-making areas and the processing of peer confidence strengthens with age.

Discussion

This study investigated how adolescents and adults process and use social information. Overall, participants across both age groups tended to rely more on their own initial judgment than on a peer's information. However, they flexibly adjusted their behavior, giving more weight to social information when they were personally uncertain or when the peer expressed high confidence. This flexible adjustment shows that people can overcome their natural tendency to favor their own information when peer input is likely to be helpful.

Adolescents and adults displayed generally similar decision-making processes, but with some key differences. Adolescents were less sensitive to changes in their own certainty and peer confidence compared to adults. Interestingly, while this study found no overall difference in social information use, in a separate, simpler task, adolescents did use more social information. This suggests that the complexity of the task or the presence of explicit reliability cues (like confidence ratings) might influence how adolescents use social information. The computational models provided a deeper understanding: they showed that adolescents behaved as if they had lower subjective certainty about both their own and a peer's information, especially in situations where they actually had more personal data. This indicates that while the behavioral outcome (amount of adjustment) might sometimes seem similar, the underlying confidence in those judgments can differ significantly.

The brain imaging results showed that many brain regions involved in processing uncertainty and social information were active in both age groups. These included areas like the prefrontal cortex, insula, and striatum. Surprisingly, adolescents showed a stronger initial neural response to their own uncertainty in some brain areas compared to adults. However, this strong initial neural response in adolescents did not translate into greater behavioral sensitivity when it came time to integrate their own and peer information. This might suggest that adolescents are still developing their ability to use these initial brain signals effectively for complex decision-making, possibly due to maturing metacognitive (self-awareness) skills.

In contrast, adults showed stronger brain activity in the ventromedial prefrontal cortex (vmPFC) when processing peer confidence. This region is critical for evaluating confidence and integrating different types of information. Furthermore, the connection between vmPFC activity and how much individuals adjusted their behavior based on peer confidence was stronger in adults than in adolescents. This suggests that the vmPFC might be less developed or active in adolescents when it comes to integrating social confidence cues into their decisions, affecting their ability to fully weigh peer information.

This study helps us understand how adolescents and adults use social information, but it has some limitations. For example, the order in which personal and peer information was presented makes it hard to pinpoint exactly when integration occurs in the brain. Future research could vary this order to see how it affects decision-making and brain activity. Also, the way "certainty" was manipulated (seeing few vs. many mushrooms) might have created a mix of different types of uncertainty, which could influence brain signals. Future studies should aim for clearer distinctions. More research with a wider range of adolescent ages and using longitudinal designs (studying the same individuals over time) could also provide a more complete picture of how these neurocognitive processes develop. Understanding how individuals learn about a peer's trustworthiness over time, rather than just being given a confidence rating, is another important area for future study.

Conclusion

In summary, this study provides insights into how adolescents and adults process and integrate social information. While both age groups use similar core decision-making strategies that combine careful consideration with simpler rules, important differences appear in how their own certainty and a peer's confidence affect their choices and brain activity. Adolescents seem less sensitive to varying levels of their own certainty and peer confidence. This reduced sensitivity might be due to ongoing development in their brain and their ability to think about their own thinking, which could affect how well they use signals of certainty when making decisions.

Introduction

Growing up, young people often pay more attention to their friends and are more likely to be swayed by them. Listening to friends can be helpful. It can teach young people how to act well and make good choices as they grow and face new things. Copying what others do can also help people learn faster, so they do not have to try and fail on their own.

However, not all advice from friends is equally good, and there are times when it is best not to listen to others. Learning when to trust what friends say is a key skill for young people to learn. How useful a friend's advice is depends on two things: what a person already knows and how much the friend knows. First, when people are in new or unclear situations and do not know what to do, they are more likely to watch what others do to guide their own actions. Young people often find themselves in new social situations, so it is not surprising that they tend to listen to their friends more. Studies have shown that young people often look for and use more information from others than adults do. Still, it is not fully clear how much young people and adults care about being unsure.

Besides a person's own level of certainty, how much the other person knows is important. In everyday life, it is often hard to tell how much someone truly knows. Instead, people often say how sure they are. Both children and adults are more likely to use information from friends who seem more confident. Even though young people generally focus more on their friends and use their advice based on how smart or popular a friend seems, how they react to a friend's confidence has not been fully studied.

Past research suggests that using information from friends is like using a special math rule where people give more weight to information that seems more certain or reliable. This means if a person is very sure about what they know from their own experience, they will trust that information more than what a friend says. On the other hand, if a friend says they are very confident, a person might trust that information more. Young people and adults also use simpler ways to decide, like copying friends if many of them do something or if friends act the same way all the time. For example, in one study, adults often stuck with their first answer even after hearing a friend's answer. It is not yet clear how these different ways of thinking change as people grow from young age to adulthood. Thinking skills often get better with age, but young people sometimes show better thinking when things are changing.

Changes in how the brain works during these years might affect how young people and adults react to being unsure and to signals from friends. Parts of the brain that help with thinking and planning are still growing in young people. This might make it harder for them to guess how sure they are or how sure their friends are. This study will compare how young people and adults use information from friends when they are unsure. Researchers predict that young people will use more information from friends and be less affected by how sure they are themselves. The study will also look at brain activity to see how different brain areas react to being unsure and to friend's confidence.

Methods

Participants

This study included 74 adults (ages 22–45) and 70 young people (ages 12–18). All participants spoke Dutch, used their right hand, had good vision, and were able to have a brain scan. Adults were going to or had finished college, and young people were going to or had finished a pre-college program. A few people were not included in the study because their information was not complete, they were outside the age range, or they did not follow the task rules. In the end, 68 adults and 64 young people were included in the behavior part of the study, and 58 adults and 47 young people were included in the brain scan part. Parents gave permission for young people under 16 to take part.

The Mushroom Game Task

Participants played a game on a computer while in the brain scanner. The game involved estimating the number of blue mushrooms in a forest. Each day, participants first looked at mushrooms in 5 nearby fields. Some days they saw many mushrooms (45 total), and other days they saw few (5 total). More mushrooms meant more information and more certainty. Participants gave their first guess using a sliding bar.

After their first guess, participants saw information from another person, described as someone who played the game before. They were first shown how confident this "friend" was (low, medium, or high confidence). After a short time, they saw the friend's actual guess. Participants could then change their own guess. The friend's confidence and guess came from real past players. Participants played 75 rounds of the game. They earned a bonus based on how close their final guess was to the real number of mushrooms. They did not get feedback after each guess.

Before the main task, participants got instructions and practiced the game. During practice, they also rated their own confidence so they could understand what the friend's confidence ratings meant. This helped make sure everyone understood the game before the brain scan.

Brain Imaging (fMRI)

Special brain scans called fMRI were done to see which parts of the brain were active during the task. Before the task, a detailed picture of each person's brain was taken. Then, during the task, a series of quick brain scans looked at blood flow, which shows brain activity. The images were prepared and cleaned up using special computer programs. Some people were not included in the brain scan study if their scans were not clear or if they moved their head too much.

How Behavior Was Measured

The main thing measured was how much people changed their first guess after seeing the friend's guess. If the number was 0, they did not change their guess at all. If it was 1, they copied the friend completely. The study used math methods to see how a person's own certainty and a friend's confidence affected how much they changed their guess. It also looked for differences between young people and adults.

To make sure any differences were not due to how well people played the game, the study also checked how far off people's first guesses were from the real answer. It also looked at how well people's own confidence ratings matched how accurate their guesses were during practice.

Understanding Decisions with Models

The study used computer models to understand how people made their decisions. These models tried to guess how people thought about their own information and the friend's information. A basic model started by assuming people used their first guess and the friend's guess to form a new, better guess.

More complex models were then tested. Some models looked at how people might think they saw more or fewer mushrooms than were actually shown, which would change how certain they felt about their own guess. Other models looked at how people might change how much they valued the friend's guess based on how confident the friend seemed.

Another important part of the model was a "stay bias." This meant that sometimes people just stuck with their first guess, even after seeing what the friend said. The models tried to figure out if this "stay bias" changed depending on how certain a person was or how confident the friend was. The best model combined these ideas: people weighed their own information and the friend's information, and sometimes they just decided to stick with their first choice. This model helped researchers see the hidden thinking steps behind why people chose to update their guesses or not.

How Brain Images Were Analyzed

Brain scan data were processed to find areas that became more active when people were more certain about their own guess or when a friend was more confident. Researchers looked for changes in brain activity at the moment a person made their first guess (when they were thinking about their own certainty) and when they saw the friend's information (when they were thinking about the friend's confidence).

The study looked at specific brain areas known to be involved in thinking about how sure one is and how others are thinking. These areas include parts of the front of the brain (PFC), a deep part of the brain called the insula, and areas involved in understanding others' thoughts (TPJ) and rewards (striatum). The goal was to see if brain responses in these areas were different for young people compared to adults.

Results

How People Used Information

People's first guesses in the mushroom game were generally close to the right answer, showing they understood the task. On average, people changed their guess by about 37% towards the friend's guess. This means they trusted their own information more than the friend's, which is common.

People changed their guess more when they were unsure themselves and when the friend seemed more confident. If a person was unsure and the friend was very confident, they might even trust the friend's information more than their own. This shows that people can change how much they rely on others when it is most helpful.

Differences Between Teens and Adults

Overall, young people and adults changed their guesses by about the same amount. However, young people were less affected by how sure they were themselves. They changed their guess less than adults when they were unsure. How much peer confidence affected their guesses was similar for both age groups.

When looking at people's reported confidence during practice, young people said they felt more confident than adults did when they were unsure. However, both groups were equally good at matching their confidence to how accurate their guesses were. This suggests that differences in how young people and adults use social information based on their own certainty might not be about how well they play the game or how they judge their own performance.

What Our Models Showed

The computer models showed that the best way to explain how people made decisions was with a model that used two main ideas: weighing personal and social information, and sometimes sticking with the first guess (a "stay bias"). This model showed that people felt more certain about their own information when there were more mushrooms to count. They also gave more weight to the friend's guess when the friend was more confident. The "stay bias" meant people were more likely to stick with their first guess when the friend was less confident.

The models also found differences between young people and adults. Young people seemed to feel less certain about their own information when they had a lot of mushrooms to count compared to adults. They also gave less weight to the friend's guess when the friend was moderately or highly confident, compared to adults. This means young people were less sensitive to how confident a friend seemed. Both groups were equally likely to use the "stay bias" and stick with their first answer.

Brain Activity and Certainty

When people made their first guess, many brain areas were active in both groups. Some areas, like parts of the front of the brain (dorsal PFC) and an area involved in rewards (caudate nucleus), showed more activity when people were more certain about their own guess. Other areas, like different parts of the front of the brain (mPFC) and areas involved in understanding others (TPJ), showed more activity when people were less certain. This shows that many brain areas work together when thinking about how sure one is.

When looking at differences between young people and adults, young people showed stronger brain activity in parts of the front and middle of the brain (anterior cingulate cortex and mPFC) when they were thinking about their own certainty. However, this difference in brain activity was not seen when they were combining their own information with the friend's information.

When seeing the friend's confidence, adults showed stronger brain activity than young people in a specific part of the front of the brain (vmPFC). This area is important for feeling confident and understanding social signals. For adults, more activity in the vmPFC was linked to being more sensitive to a friend's confidence, but this link was not found in young people. This suggests that the connection between brain activity and how much people care about a friend's confidence gets stronger with age.

Discussion

This study looked at how young people and adults use information from others and what happens in their brains. In general, people of all ages tended to trust their own ideas more than what a friend said. But they changed their minds more when they were unsure or when a friend seemed very confident. Young people and adults thought in similar ways, combining their own information with a friend's, but also sometimes just sticking with their first idea.

However, there were important differences. Young people were less affected by how sure they were themselves and by how confident a friend seemed. This might be because their brains and thinking skills are still growing. For example, young people's brains showed a stronger response when they were first unsure, but this strong response did not seem to help them as much when they later combined their own and a friend's information. Adults, on the other hand, showed a stronger brain response in an area called the vmPFC when a friend seemed more confident. This brain area is important for understanding confidence and social cues. This could mean that adults are better at using a friend's confidence to make decisions.

This study helps us understand how young people and adults use information from others when they are not completely sure. However, the study had some limits. For example, the way the information was shown might have made it hard to fully separate how people thought about their own certainty versus a friend's. Future studies could change the order of information or how it is shown to learn even more. Also, it would be good to study a wider range of young ages over time to see exactly how these brain and thinking skills develop.

Conclusion

In short, this study adds to our knowledge of how young people and adults process and combine information from others. While both groups use similar basic ways of thinking, how much their own certainty and a friend's confidence affect their actions and brain activity is different. Young people seem to be less sensitive to how sure they are and how confident a friend is. This may be because their brains and thinking abilities are still developing, which could make it harder for them to use these certainty signals well.