Friend or Foe: Delegating to an AI whose Alignment is Unknown
2026
Accepted for presentation at EC'26
Abstract
We study delegation of a risky decision to an AI whose objective may be aligned or misaligned with the designer's. The designer limits the AI's discretion using treatment-rate constraints and restrictions on input informativeness. We characterize the best- and worst-case payoff frontier and the choices that implement it. As the designer accepts more downside risk, delegation expands from individuals with more uncertain outcomes to individuals with more certain ones. Conditional on delegation, the optimal input restrictions are asymmetric, permitting informativeness in one direction (e.g., treatment success) but restricting it in the other (e.g., treatment harm).
BibTeX
@unpublished{fudenberg2026friend,
author = {Drew Fudenberg and Annie Liang},
title = {Friend or Foe: Delegating to an {AI} whose Alignment is Unknown},
year = {2026},
note = {Working paper}
}