Bounding the Long Tail: Ai Norms for Decision-Making Under Negligible Probabilities

The issue discussed in this paper is a long-standing issue of decision theory. It is reframed as a design problem for intelligent agents. It provides a principled cutoff which can be ultra-low-probability and extreme utility outcomes. This cutoff can prevent the exploitability of autonomous agents. A vulnerability class is characterised for expected-utility maximisers and a rationally negligible probability threshold, which is rooted in a scepticism standpoint, is introduced. This preserves dominance and tractability and blocks adversarial gambles (Pascal-type offers). The use of formal analysis motivates the design norms for AI agents (utility bounding, calibrated priors and epsilonscreening) and guidance on the selection of context-sensitive thresholds that ensure that preferences do not undergo dramatic changes. The proposition of a safety-oriented inductive bias for rational AI decision makers, whose theorists' desiderata are in line with implementable policy constraints in high stakes low signal situations.