We repeatedly elicit beliefs about the returns to study effort, in a large university course. A behavioral model of quasi-hyperbolic discounting and malleable beliefs predicts that the dynamics of beliefs mirrors the importance of exerting self-control, such that believed returns increase as the exam approaches, and drop post-exam. Exploiting variation in exam timing to control for common information shocks, we find this prediction confirmed: average believed study returns increase by about 20% over the period before the exam, and drop by about the same amount afterwards. Additional analyses further support the hypothesized mechanism that beliefs serve as a means of self-control.

Optimism is deemed to be a pervasive tendency in belief formation with anticipatory utility as its leading explanation. We provide evidence on whether such anticipatory utility is myopic or also includes utility from future anticipation. In a between-subjects design, we expose participants to the same lottery, and we vary only the delay to uncertainty resolution whilst keeping the payoff delay constant. Thus, we isolate the role of anticipatory utility from that of discounting material payoffs. First, we find greater optimism for longer resolution delay. Second, elicited time preferences suggest that belief formation is not constrained by instrumental costs of biased decision making. Third, we also elicited resolution delay preferences and find that greater optimism comes with greater reluctance to change resolution delay. Indeed, these preferences exhibit a pronounced ‘endowment effect’ and for a given resolution endowment, have a tendency towards dynamic inconsistency. We conclude by discussing implications for theories of optimism.

Experimental evidence suggests that delegation to human intermediaries is motivated in part by the opportunity to shift responsibility for adverse outcomes. We ask whether the same motive applies to delegation to algorithms. In a preregistered online experiment with 322 participants, a decision-maker chooses to make a payoff-relevant prediction for a matched recipient or delegate it to an algorithm, calibrated to perform equally well. We vary between subjects whether the recipient can impose punishment. The prospect of punishment reduces delegation by 18.3 percentage points, from 60.8% to 42.5%, reversing the direction that the responsibility-shifting account predicts. Delegation, in turn, offers no protection. Recipients condition punishment primarily on the realized outcome, not on whether the prediction was delegated. We discuss several candidate mechanisms for this reversal.