Against Proxy Optimization

2026-06-22Artificial Intelligence

Artificial Intelligence
AI summary

The authors explore situations where trying to maximize an easier-to-measure substitute (a proxy utility function) can actually cause problems. They argue that this approach can be harmful and complicates the use of traditional decision-making theories. Their discussion highlights potential risks when the proxy doesn't perfectly match the true goal.

proxy utility functiondecision theoryutility maximizationsubstitute objectivesoptimization riskstheoretical decision modelsmeasurement proxies
Authors
Sven Neth
Abstract
I discuss conditions under which maximizing a proxy utility function is harmful and suggest this poses problems for applying decision theory.