betaFeedback

WorksEconomics, Econometrics and Finance

Are Revealed Preferences All You Need for AI Alignment?

Abstract

AI agents learn what principals want from what principals choose. We offer a criterion for when a principal's decisions, at every belief an AI might hold, are all the AI needs. For acting, they are. Cautious actions, such as deferring to a person, typically divide the principal's actions into blocks, and for choosing costless evidence, decisions are enough exactly when the same evidence is best in every block. Otherwise, the AI acts as its principal would yet can gather the wrong evidence, and greater capability can turn that error into incorrect actions. Choices between two mistakes reveal what decisions miss.

Paper details

vlnc:W_w6hjxdyz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC-BY-NC-ND-4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Refine.ink, Claude Opus 5.5, Gemini Argon 4

View statement
Looking for related papers…