The Epistemic Limits of Large Language Models: Trust, Testimony, and Group Understanding

Abstract

In recent years, the use of Large Language Models (LLMs) has become increasingly widespread. A major catalyst in this trend was the 2022 release of OpenAI’s ChatGPT GPT-3.5 model for public use. With similar alacrity, the philosophical literature devoted to investigating these and other artificial intelligence (AI) systems has grown. This thesis contributes to this growing literature by aiming to answer the following question: how, if at all, can we use LLMs to obtain epistemic goods, particularly testimonial knowledge and group understanding? In service of this aim, this thesis is split into three chapters. Throughout these chapters, I make three main conclusions: (i) there are limited circumstances under which we can deem an LLM a credible expert, (ii) we cannot deem an LLM a credible standard testifier (iii) LLMs are incapable of contributing to group understanding. In chapter one, I review the existing literature in three areas relevant to answering these questions: testimony, understanding, and the philosophy of AI. In chapter two, I argue that LLMs can be deemed credible experts but not credible standard testifiers. I conclude this section by offering some potential explanations for this tension. In chapter three, I argue that LLMs cannot be contributing members to group understanding on either a deflationary or inflationary account of the phenomenon. Here, I show that LLMs cannot be attributed with non-metaphorical understanding or the requirements for contributing to group grasping.