Artificial Intelligence as a Source of Risk in Asset Management

Abstract

Artificial Intelligence promises investors more information, since it can read more documents and process more text than any human. This article argues that the opposite might happen. Once AI outputs matter for investment decisions, the parties being measured learn what the machine rewards and write for it, trading a heterogeneous set of human readers for one predictable target. Automation strips away the dispersed and skeptical, though bounded, scrutiny that once made it harder to trick readers, leaving investors with more data and less knowledge. This risk, termed technological epistemic risk, is related to but distinct from statistical model risk: The model is not wrong, but its deployment changes what it reads. The article derives one practical rule: Rank extracted signals by both how costly they are to fake and how well they have predicted outcomes, trusting only signals that score well on both.