On the Beliefs of Large Language Models
Abstract
Do large language models (LLMs) have beliefs? I argue that they do, but that their beliefs represent a degenerate case of the phenomenon. The argument turns on Devin Curry’s distinction between the attitude of belief and the intrinsic cognitive state of belief. Focusing exclusively on the attitude, construed here as a behavioral phenomenon, I argue that ascribing beliefs to language models offers greater predictive utility than alternative methods of predicting their outputs, and that this predictive success licenses a form of realism about LLM belief. I then argue that LLM belief suffers from a lack of integrity. The transformer architecture, on which standard LLMs are based, has no intrinsic mechanism for carrying information across context windows. Moreover, because LLM behavioral dispositions are prompt-sensitive, they are prone to self-contradiction across contexts. Humans contradict themselves as well, but our rational integrity is maintained by a suite of mechanisms that large language models lack. The paper concludes by asking whether the degenerate nature of LLM belief demands that we change our current folk psychological habits. The answer is that change is demanded, but it is a change in the construal behind the words we choose, rather than in the words themselves. In order to speak and reason accurately about the doxastic states of LLMs, a novel construal will have to be learned.
