Bounding the Long Tail: AI Norms for Decision-Making under Negligible Probabilities

This paper recasts a long-standing decision-theory dilemma as an AI agent-design problem, proposing a principled cutoff for ultra-low-probability, extreme-utility outcomes to prevent exploitability in autonomous systems. By characterizing a vulnerability class for expected-utility maximizers and introducing a rationally negligible probability threshold grounded in cognitive skepticism, the framework preserves dominance and tractability while blocking adversarial gambles such as Pascal-type offers. Formal analysis motivates design norms for AI agents—utility bounding, calibrated priors, and epsilon-screening—together with guidance on selecting context-sensitive thresholds to maintain preference stability. This positions the proposal as a safety-centric inductive bias for rational AI decision-makers, aligning theoretical desiderata with implementable policy constraints in high-stakes, low-signal environments