Do Large Language Models Believe?

Abstract

Here is an increasingly popular view: if believing that P is just to be disposed to assert, infer, and act as if you take P to be true, then large language models (LLMs) have beliefs. On this view, LLMs pass the relevant behavioral tests: they assert truths, pass theory-of-mind evaluations, and exhibit a degree of internal consistency. This talk pushes back. I show that even if we accept the dispositionalist conception of belief, current LLMs do not meet relevant standards for an important kind of belief ascription. I develop three arguments for this claim. First, the coherence threshold: belief attribution presupposes some degree of internal consistency. But LLMs are so prompt-sensitive that their belief-like assertions are subject to radical incoherence even within one context window. There is no practice-independent fact about how much logical coherence belief-ascription requires. So we are free to demand a level of coherence high enough to support explanation and prediction. Given the volatility of LLM assertion, we can easily choose to impose a coherence threshold that they cannot yet meet. Second, the proportionality constraint: in humans and non-human animals, belief ascription is only predictively useful when paired with a model of desire. In LLMs, inconsistency carries no cost, because nothing is at stake. Since nothing depends on which belief-like attitudes the system adopts, nothing stabilizes its content-bearing mechanisms over time. Third, the mindshaping argument: in humans, attitudes are aligned—imperfectly but meaningfully—with cognitive mechanisms, due to social pressures that make us accountable for what we say and do. This alignment makes belief-ascription explanatory rather than merely predictive. In LLMs, this alignment mechanism is absent. Their norm-sensitive belief-like assertions are not anchored to underlying cognitive mechanisms at inference time. Consequently, belief-ascription risks tracking surface regularities that fail to generalize. These three arguments show that belief-ascription to LLMs faces philosophical obstacles. I wrap up by noting that the situation may look different for the emerging class of AI agents — systems with persistent memory, planning capacity, and goal-conditioned behavior.

Charles Rathkopf

I am interested in how mental properties emerge from physical stuff.