CS Colloquium

Time and Location:

Speaker:

Link:

Modern AI systems interact repeatedly with humans and with each other, shaping marketplaces and downstream decisions. I study learning and strategic interaction in these human-AI systems, with the goal of understanding when these interactions lead to complementarity and alignment, and when they instead produce collusion or instability. In this talk I will focus on two threads. First, I will present models of human-AI complementarity, showing how humans and AI systems with different information can, through repeated interaction, reach accurate, consistent predictions without a shared prior, while guaranteeing that good-faith participants are never harmed by collaborating. Second, I will describe an AI selection game in which a human chooses among partially misaligned AI advisors, and show that under a natural “market alignment” condition, competition among advisors guarantees outcomes close to those of a perfectly aligned assistant. I will briefly discuss how these results connect to my work on algorithmic collusion in markets and on “menus” of algorithms that make the long-run commitments of learning systems explicit and robust.