Arata Matsuda

This paper explores the ways in which large language models (LLMs) can be used for the purpose of experimental philosophy. LLMs can be used in different ways for different purposes in experimental philosophy. Among others, we focus on what we call the “non-predictive” research, in which LLMs’ “intuitive” responses to thought experiment cases are not expected to be predictive of, or analogous to, human subjects’ intuitive responses to them. In effect, the non-predictive research regards LLMs’ responses as superior to human subjects’ responses, at least for the purpose of experimental philosophy. We argue that the non-predictive use of LLMs in experimental philosophy is highly motivated. Firstly, human subjects’ intuitive responses tend to be vulnerable to irrelevant factors and conceptual misunderstandings, which raises a serious worry concerning the reliability of their responses. Secondly, it is possible in the near future that LLMs can overcome these problematic factors to which human subjects are subject. As a case study, we conducted an experimental study in which we examined LLM’s “intuitions” on free will and determinism. Our mixed results suggest that while current LLMs outperform human participants in terms of comprehension—especially in understanding deterministic scenarios—they remain highly sensitive to irrelevant factors such as framing and question order. These findings indicate that although LLMs have promising potential, they are not yet reliable sources of stable and unbiased philosophical intuitions. We conclude that substantial technical improvements are required before artificial intuitions can serve as superior alternatives to human intuitions in philosophical research.

Philosophers commonly use thought experiments to elicit intuitions that serve as evidence for or against philosophical theories. Daniel Dennett calls such thought experiments “intuition pumps” and expresses concern that some of them are “deceptive” in that they elicit conflicting intuitions and thus lead to conflicting conclusions. What, then, should philosophers do with “deceptive” intuition pumps? Dennett proposes what might be called restrictionism, according to which philosophers should restrict the use of deceptive intuition pumps. In this paper, however, we argue that it faces what we call the problem of over-restriction: if philosophers restrict the use of deceptive intuition pumps, then they end up restricting the use of so many intuition pumps that they cannot continue philosophical inquiry. Then, as an alternative, we propose reformism, according to which philosophers should continue the use of deceptive intuition pumps after reforming them. We argue that reformism is more attractive than restrictionism, because it deals with the deceptiveness of intuition pumps while avoiding the problem of over-restriction. Overall, through criticizing Dennett’s restrictionism, we aim to illuminate the attractiveness of reformism, which has been underappreciated as a response to the deceptiveness of intuition pumps, and to call for future research in the reformist research program.