Computational Modeling shows Confirmation Bias during Formation and Revision of Self-Beliefs

Abstract Self-belief formation and revision strongly depend on social feedback. Accordingly, self-beliefs are subject to (re)evaluation and updating when facing new information. However, it has been shown that self-related learning is rarely purely information-driven. Instead, self-related learning is susceptible to a wide variety of biases. Among them is the confirmation bias, which can render updating insufficient, leading to inaccurate self-beliefs. To better understand these biases, it is important to delineate the effects of initial expectations towards the self and the confidence associated with the self-belief. In a novel behavioral approach, we introduced two learning phases during which participants completed an estimation task and received feedback allegedly related to their performance. In the first session (T1) , participants established beliefs about their abilities in this task based on trial-by-trial feedback. In the second session (T2) , participants received feedback that differed substantially from the feedback they had received at T1 , thus creating the possibility for belief revision. Computational modeling was used to describe initial belief formation and later revision. The results showed confirmatory belief updating behavior on different levels: Participants did not, on average, revise their beliefs at T2 , although they were constantly confronted with conflicting evidence. Instead, we observed that initial expectations were linked to biased learning from the received feedback, even at the beginning of the initial belief formation phase. Further, higher confidence in the beliefs was associated with attenuated revision. Together, the results underline the importance of individual priors when delineating learning biases.