IHD Colloquium 9/11/23 Theoretical and Applied Implications of Large Language Models for Developmental Science
September 11, 2023 • 12:10pm–1:30pm • 1102 Berkeley Way West (in person talk)
- Syntax and Semantics in the Age of Large Language Models
Steve Piantadosi
The recent rise in large language models profoundly changes the landscape for theories of human language. I'll discuss how these models should cause us to rethink many popular ideas about language acquisition, including most prominently those argued for by Chomsky. I'll also discuss the way in which these models implement theories of language and grammar, as well as the links and gaps between these models and child language learning. Despite important differences, I'll argue that people who care about learning should take LLMs seriously.
- How to build AI platforms and policies that promote—instead of impede—children’s curiosity and learning
Celeste Kidd & Bill Thompson, with contributions from Evan Orticio and Alex Yang
Large language models (LLMs) are being adopted at a breakneck pace, and children are at the forefront. A majority of kids ages 12-18 use ChatGPT, while their parents lag behind (Common Sense Media, 2023). Despite this, we do not know how children use LLMs, how they conceptualize them, and how their intellectual character and beliefs are shaped by them. Past work suggests LLMs’ confident, agentive outputs diminish curiosity in users—especially children—leaving them vulnerable to adopting fabrications as established beliefs (Kidd & Birhane, Science, 2023). We propose addressing this by constructing a platform for studying children’s LLM use, how LLMs influence children’s character, and how LLMs could be redesigned to promote character development and accurate beliefs. We will discuss new ideas for how we can conduct research to inform the development of a new generation of LLMs—and science-informed policies and educational practices that support children’s character development, promote curiosity, and promote accurate belief adoption. This work is in the very early stages of development—which is happening in conversation with graduate students Evan Orticio and Alex Yang—so we’ll be briefly presenting our basic premise and ideas for empirical studies, and eliciting feedback from all of you about which of many possible avenues you think are most fruitful.