Transmission Versus Truth: AI Models as Cultural Technologies and Epistemic Agents

Join us for a talk by Alison Gopnik, Distinguished Professor of Psychology, Affiliate Professor of Philosophy and member of the Berkeley Artificial Intelligence Research Group. This talk is part of the Kempner Seminar Series, a research-level seminar series on recent advances in the field.

Abstract: Many approaches to AI assume that there is a single “general intelligence” which AI systems should maximize. I argue instead that there are multiple conflicting intelligences which trade-off against each other. In particular, there is a contrast between discovering the truth about the world and transmitting that information to other agents. LLM’s are extremely effective cultural technologies, analogous to writing, print or internet search. They summarize and transmit information that humans have already learned but they are not designed to recover the structure of the external world. In contrast, children are extremely effective truth-seeking agents, exploring and experimenting. I suggest that recent intrinsically motivated reinforcement learning agents are a  better model for the truth-seeking we see in children. In particular, I argue that RL systems that use “empowerment gain” as an intrinsic reward, seeking actions that lead to predictable outcomes, might be an important mechanism for the causal learning that is key to children’s everyday theory formation. I will report some recent empirical work along these lines comparing how children and AI agents explore in virtual environments.