Insensitivity to how our actions affect others drives inflated self-evaluations during interactions which involve both shared and conflicting goals.

Receiving affirmation, whether from ourselves or others, is crucial for interpersonal relationships and goes awry in mental disorders. Meaningful evaluations emerge during interactions, where people can support or let each other down. However, such dynamics are difficult to understand comprehensively without quantitative theory and modeling. Here, we implemented an interactive decision-making game wherein two real-life participants evaluated, i.e. graded their approval, for themselves and their play partner. Young adult participants interacted in a multi-level version of the iterated prisoner’s dilemma. Crucially, each participant did not interact with the other directly, but instructed an avatar to do so on their behalf. This allowed increased experimental control while preserving considerable ecological validity. We tested computational models of participants’ evaluations of self and other, based on their beliefs about the quality of their decisions. However such models were less successful than a novel class of models, where self- and other- evaluations depended directly on the combination of self- and other- outcomes. The winning models suggested that for a given participant, evaluation of the self is proportional to how much one’s partner benefits, and vice versa. We found marked self-positivity bias, especially in dyads where neither partner cooperated. This was consistent with attributional theory, negatively evaluating others rather than the self for adverse outcomes. Between participants, self-positivity bias was explained by a reduced weight of one’s partner’s benefits for self-evaluation, hinting that the negative outcome subject to external attribution was the partner’s, rather than one’s own, poor returns. Preliminary analysis also suggested that a reduced sensitivity to others’ outcomes was associated, in this context when participants may have both cooperative and competitive motives, with reduced earnings for the self. The proposed computational model provides a concise and novel account of self-serving bias in evaluations, clearly observed during interactions.