Making good decisions is essential for survival and success, yet humans and animals often exhibit perplexing irrational decision-making whose biological origin remains poorly understood. Recent empirical and computational work suggests that altered computations in perceptual, motor and memory systems in the brain may arise from informational, metabolic or robustness constraints on their internal connectivity structure. However, whether and how such neurobiological constraints may have molded the architecture of decision systems (such as the orbitofrontal cortex) and eventually distorted decision-relevant computations, remains largely unknown. We first train cohorts of artificial neural nets to perform ten variants of rational decision-relevant computations. Those variants that operate under a specific option-encoding format exhibit most of the electrophysiological coding properties observed in orbitofrontal neurons of monkeys making decisions under risk. We then distort these neural nets’ internal wiring to reproduce monkeys’ irrational choices. This induces deterministic spillover interferences in decision-relevant computations that generalize across individuals, at both the behavioral and neural level. Importantly, although irrational nets do not seem to bring informational or metabolic benefits, they display enhanced tolerance to damage and noise when compared to their rational counterparts. This suggests that some forms of irrational behavior may be the incidental outcome of distal evolutionary pressure on the robustness of orbitofrontal circuits.