Living with Artificial Agents

Peter Railton (University of Michigan, Ann Arbor)

A lecture, free and open to the public with a reception to follow.

Abstract

"In order to overcome the limitations of Large Language Models trained exclusively on linguistic data, AI development is now widening and deepening AI systems' engagement with the world through agency--behavior that is internally-guided, in the sense that it is based upon the systems' own perception of the environment and ability to use this perceptual information to select actions in light of internal goals.  However, along with the promise of improved performance, this brings with it the risk that artificial agents will escape human control.  Some form of self-regulation responsive to concerns of safety and ethics needs to be part of AI agents' internal guidance--just as it must be if human agents are to behave morally.  Interestingly, the psychology of the human development of autonomous moral capacities increasingly emphasizes the role of experiential learning over innate principles--might this enable us to see how artificial agents, whose great strength is learning, could acquire sensitivity to concerns of safety and ethics?."