Nina Van Rooy

Recently, several researchers have suggested that we should attribute theory of mind (ToM) to large language models (LLMs) such as GPT-4, because they pass tests designed to measure ToM – most prominently the false-belief test, a paradigmatic measure of ToM in humans and other animals. In this paper I argue that LLMs passing ToM tasks provides only weak evidence for the claim that they have ToM. Success on ToM tests is not by itself sufficient to warrant attribution of ToM: substantive additional evidence is needed to demonstrate that LLMs have several empirically plausible prerequisites for ToM ability, and to rule out alternative explanations for their performance on ToM tasks. I draw on the evolutionary, developmental, and comparative psychology of ToM to make two main points. First, I defend a general principle that operationalizations of cognitive constructs should be tailored to the kind of agent they are being applied to, and thus that we cannot take for granted that ToM measures designed for humans or other animals remain construct valid (if they were construct valid at all to begin with) in the case of LLMs. More precisely, I argue that there are several notable prerequisites for ToM possession, and thus also for meaningful ToM testing, that likely are not met in LLMs. Second, I demonstrate that ToM attribution to non-human animals has not been nearly as simple as achieving successful performance on the false-belief test: competing hypotheses that did not involve mindreading were repeatedly offered as explanations of animals’ performance on ToM tasks. Thus, before ToM attribution was warranted, comparative psychologists had to conceive of and then rule out alternative plausible hypotheses, a process that took several decades. These methodological lessons from the psychology of ToM are important for AI researchers to heed.

In The Entangled Brain, Pessoa criticizes standard approaches in cognitive neuroscience in which the brain is seen as a functionally decomposable, modular system with causal operations built up hierarchically. Instead, he advocates for an emergentist perspective whereby dynamic brain networks are associated, not with traditional psychological categories, but with behavioral functions characterized in evolutionary terms. Here, we raise a number of concerns with such a radical approach. We ultimately believe that while much revision to cognitive neuroscience is welcome and needed, Pessoa’s more radical proposals may be counterproductive.