How not to argue against AI thought

This paper identifies and criticizes a common line of argument for the negative conclusion that Large Language Models (LLMs) cannot think, or for the closely related conclusions that they cannot understand or that their outputs are meaningless. This line of argument can be found in several recent papers. We begin by discussing a representative example – Stoljar, D., and Z. Zhang. 2026. “Why ChatGPT doesn’t Think: An Argument from Rationality.” Inquiry 69(6): 3011–3039 – and present three objections to it: first, it renders the successful performance of LLMs on a wide range of tasks miraculous; second, it does not give sufficient weight to the possibility that LLMs could be extracting information about the world from patterns of word co-occurrence; and third, it overgeneralizes, leading to the conclusion that human inferences could only ever be based on premises about proximal stimulus patterns in the brain. We then discuss recent arguments by Hattiangadi, A., and A. Schoubye. forthcoming. “The Outputs of Large Language Models are Meaningless.” In Communicating with AI: Philosophical Perspectives, edited by H. Cappelen and R. Sterken. Oxford University Press and Titus, L. 2024. “Does ChatGPT have Semantic Understanding? A Problem with the Statistics-Of-Occurrence Strategy.” Cognitive Systems Research 83 (101174): 1–13, explain how our objections extend to them and identify the underlying reason that makes these sorts of arguments tempting. We finish by drawing a general lesson for debates surrounding thinking in LLMs.