The task implemented key features of a realistic social learning situation, in which the learner receives well-intentioned advice prior to individual learning, but good advice is not certain. (A) In the learning task, participants repeatedly chose from one of four card decks associated with different gains and losses (c.f. ), with the goal of maximizing cumulative rewards. Each trial started with the presentation of the four decks. Participants had 2.5 s to choose a deck. After a variable fixation interval, feedback was presented for 2.5 s. (B) Advisors performed the 4-armed bandit task in the laboratory and indicated their advice on a form, which advice-receivers obtained before performing the task in the scanner.

(A) Participants' average choice proportions (running average of 11 trials ±2 standard errors) for the recommended and the non-recommended corresponding deck. This sustained influence of advice is consistent with the outcome-bonus mechanism, but not with the assumption that advice influences only the evaluation of choice options prior to individual learning (see below for detailed description of the social learning models). (B) The panel shows Bayes factors comparing the two models with the lowest Bayesian information criterion (outcome-bonus model and prior+outcome-bonus model) with the other models. The Bayes factor shows how much more likely the two best models are as compared to the model in the respective row. For instance, the last row shows that the prior+outcome-bonus model is about 30 times and the outcome-bonus model is about 40 times more likely than the individual learning model. Models assuming the influence of advice on evaluation of either gains or losses performed worse than the original outcome-bonus model. Moreover, more complex models implementing a dynamic outcome-bonus do not predict the data better than the simpler models with stable outcome-bonus. The inset (same order of models on the x-axis) shows that these results also hold when using the Akaike information criterion (AIC) as a model selection criterion. Because the AIC imposes smaller penalties for additional free parameters (for our sample size), models with a dynamic outcome-bonus perform better according to the AIC criterion. In sum, regardless of the model selection criterion, an outcome-bonus and a prior mechanism are implemented in the best models, but a dynamic or payoff-selective outcome-bonus mechanism is not.

We used computer simulations to compare the average rewards gained by the prior model and the outcome-bonus model when performing a 4-armed bandit task with 100 trials after receiving advice about which bandit has the highest payoffs. To examine the models' performance over a range of learning settings, simulations were repeated for different task difficulties, levels of social influence, probabilities of correct advice, and learning parameters (see ). (A) Difference in average payoffs across learning environments for 50% good advice. Each cell depicts the difference in average payoffs of the two models for a particular combination of task difficulty and social influence. The difference was computed from 1,000 simulated learning tasks for each model, whereby each model received good advice in half of the tasks. Difficulty, defined as difference in mean payoffs divided by payoff variance, is varied on the y-axis. Social influence, defined as either the magnitude of outcome-bonus (β_b) or higher initial reward expectation for the recommended option (β_p), is varied on the x-axis. The outcome-bonus model is more likely to choose a highly rewarded option (and thus accumulates most reward) for most levels of task difficulty and social influence, with exception of the combination of bad advice in a difficult task with strong social influence. (B) Learning curves show that the outcome-bonus model is better because it profits from good advice in the long run, even if individual learning already leads to a clear preference for a good bandit. The inset depicts cumulative rewards for both models. After bad advice, cumulative payoffs are reduced less for the outcome-bonus model compared to the prior model (see inset). In contrast, the prior model does not profit from good advice in the long run, and cumulative payoffs are greatly reduced after bad advice (see also –). The inset also highlights that bad advice harms the prior model because it abolishes choices of the better option until the prior expectation induced by advice has been unlearned.

(A) Feedback after following advice led to greater activity in the left caudate head and septal area. (B) Parameter estimates (PE) of the General Linear Model analysis suggest that the outcome-bonus is implemented in the septal area and left caudate as a greater increase in BOLD signal for gains and as a smaller decrease in BOLD signal for losses after following advice. (C and D) VST, VMPFC, and OFC showed a greater reduction in BOLD response to negative outcomes if these followed deviation from advice. (E) Participants with higher outcome-bonus parameters showed greater increases in BOLD signal in the left lateral amygdala for gains after following advice compared to after not following advice. (F) Scatter plot illustrating this correlation (1 outlier participant removed). Note that all regions implementing the outcome-bonus are gain preferring after not following advice, in that they all show an increase in BOLD signal to gains and a decrease in BOLD signal to losses.

(A) Regions showing greater BOLD response to positive than to negative feedback after not following advice. Top left: VMPFC (max. z-score = 3.50, x = −4, y = 50, z = −18); top right: Left VST (max z-score  = 3.73, x = −16, y = 12, z = −10); bottom left: Insular cortex (max z-score  = 3.33, x = 38, y = 8, z = −16). (B) Estimated parameter estimates show that following advice led to an attenuation of the standard reward signal (as identified from trials in which participants did not follow advice) in the VMPFC and the VST.