g., see Figure S5). Subjects made decisions in three types of trials (see Figure 1). In condition 1, they were presented with a face picture of a human agent and had to decide whether to bet Capmatinib manufacturer for or against the agent. After a brief delay, they observed the agent’s prediction about the asset performance (up/down). Following a jittered interstimulus interval, feedback was presented indicating whether the asset’s value went up or down, as well as feedback for the trial. The subject made $1 if she guessed correctly the performance of the agent (i.e., if she bet for him, and he was correct, or if she bet against him, and he was mistaken) and lost $1 otherwise. This screen also indicated the performance

of the asset with an up/down arrow, independently of any other contingencies for the trial. The feedback phase was followed by

a jittered intertrial interval. Condition 2 was identical to condition 1, except that now the agent was depicted by a 2D fractal image and described to the subjects as a computerized-choice algorithm. In contrast, in condition 1, the agent was described as depicting the predictions of a real person that had made predictions in a prior testing session. This was indeed the procedure implemented, although the choices that the real person made in the prior testing session were predetermined by choices generated by the probabilities shown in Figure 2A. In condition 3, there was no agent and thus no ability prediction. Instead, the subject had to predict whether the asset would go up or down. The participant’s payoff in this case depended on the ability

OSI-906 chemical structure to predict the next outcome Tolmetin of the asset correctly: $1 for correct guesses, and −$1 for incorrect ones. We emphasize that in all of the conditions, the subject’s payoff depended on the quality of his guesses, and not on the actual performance of the asset or of the agents. At the end of the experiment, subjects were paid their total earnings in cash. The task was divided into four fMRI blocks (or runs) of 55 trials. In each block, the subject observed the predictions of three agents (either two people and one algorithm, or the reverse). There were 11 asset prediction trials per block. Subjects made predictions about each of the three agents in a block in an equal or nearly equal number of trials (14 or 15 trials each, depending on the block). The three agents and asset prediction trials were randomly interleaved with the constraint that the same stimulus (agent or asset) was never repeated. In total, this allowed for 88 trials observing people, 88 trials observing algorithms, and 44 asset prediction trials. There were four people and four algorithms in total. Each agent was characterized by a fixed ability α denoting the constant and independent probability with which he made the correct prediction for the asset’s performance in every trial.