Speaker’s Use of Mental Verbs to Convey Belief States: A Comparison between Humans and Large Language Model (LLM)

While LLMs have good knowledge of linguistic rules and regularities, whether they possess functional linguistic competence, or the ability to use language in real-world settings, remains contested (Mahowald et al., 2024). Pragmatics is considered to be part of functional linguistic competence, since it requires social reasoning abilities beyond pure linguistic knowledge. Some studies found that LLMs exhibit pragmatic abilities comparable to humans (Chen et al., 2024; Hu et al., 2023), while others observed potential deficiencies (Qiu et al., 2023).
The current study focused on speaker’s use of different mental verbs to convey the beliefs of themselves and others. For instance, if the speaker says Sean knows that the book is in the box, it likely suggests that both the speaker and the attitude holder (i.e., Sean) believes that the book is in the box although the utterance does not explicitly discuss the speaker’s belief. Pan and Degen (2023) examined this phenomenon using the Rational Speech Act framework, where speakers choose an utterance according to its informativeness and cost, while listeners interpret an utterance based on the probability for the speaker to produce it given different possible meanings. Interestingly, they found an anti-veridical effect for utterances using think, although think is considered to indicate uncertainty about the embedded content. A similar phenomenon was observed in GPT, which was more likely to infer that the attitude holder had a false belief when prompted with think (Trott et al., 2023). We speculated that the anti-veridical effect of think was because listeners would expect know to be used instead when the speaker believes in the embedded content. To explore this possibility, we investigated how humans and GPT-4 use know, think, and unembedded bare forms given different belief states of the speaker and attitude holder. We hypothesized that humans would tend to use more informative utterances, while GPT-4’s responses may not fully account for human behavior due to limited functional linguistic competence.
Our analyses indicated that GPT-4 indeed could not fully account for human responses. However, humans and GPT-4 both showed higher preferences for the bare form than hypothesized, despite it being less informative in some cases. This result could be because the bare form is shorter than the other forms, thus having lower utterance cost. Meanwhile, humans also preferred think over know, even when know was more informative, but GPT-4 did not. Although they have equal lengths, know may be more costly because it requires information about both the attitude holder’s belief and the truth of the embedded content. GPT-4’s inability to recognize the higher cost of know may be because LLMs do not form world models like humans do (Yildirim & Paul, 2024) and thus may select utterances based on statistical regularities without considering the cost involved in communication. Finally, despite humans’ general preference for think over know, know is still more prevalent when the speaker also believes in the embedded content than under all other situations, suggesting that pragmatic competition between know and think may still explain the anti-veridical effect of think.