Defining Knowledge: Bridging Epistemology and Large Language Models
Constanza Fierro, Ruchira Dhar, Filippos Stamatiou, Nicolas Garneau, Anders Søgaard
TL;DR
The paper tackles the question of what it means for LLMs to know and demonstrates that NLP linguistics often treats knowledge without a solid epistemological footing. It surveys five classical definitions of knowledge—tb-knowledge, j-knowledge, g-knowledge, v-knowledge, and p-knowledge—and formalizes their mappings to LLM assessment, then compares these mappings to current NLP evaluation practices. An empirical survey of 105 philosophers and computer scientists reveals meaningful disagreements across definitions and an overall trend that non-human knowledge is possible while empirical knowledge in LLMs is contested or debated. The authors propose concrete, definition-aligned evaluation protocols and argue that grounding knowledge claims in epistemology can lead to more rigorous, trustworthy assessments of what LLMs truly know and how to test it.
Abstract
Knowledge claims are abundant in the literature on large language models (LLMs); but can we say that GPT-4 truly "knows" the Earth is round? To address this question, we review standard definitions of knowledge in epistemology and we formalize interpretations applicable to LLMs. In doing so, we identify inconsistencies and gaps in how current NLP research conceptualizes knowledge with respect to epistemological frameworks. Additionally, we conduct a survey of 100 professional philosophers and computer scientists to compare their preferences in knowledge definitions and their views on whether LLMs can really be said to know. Finally, we suggest evaluation protocols for testing knowledge in accordance to the most relevant definitions.