LLMs Struggle to Reject False Presuppositions when Misinformation Stakes are High

These implicit assumptions, known as presuppositions, refer to background knowledge or shared beliefs assumed to be part of the common ground between interlocutors (Stalnaker, 1973). Presuppositions are introduced by specific words or syntactic structures called presupposition triggers. For instance, in the statement, “As you know, we’ve completed the wall,” the factive verb ’know’ triggers the presupposition that it is true that the wall has been completed. Presupposition triggers include a wide range of elements such as factive verbs, iteratives, or definite descriptions Levinson (1983). They play an integral role in communication, allowing speakers to rely on shared knowledge without making it explicit. However, presuppositions are not limited to reaffirming shared beliefs; they can also introduce new information into the common ground, a phenomenon known as informative presuppositions (Tonhauser, 2015). Such presuppositions are particularly prominent in persuasive contexts, including media and political discourse (Sbis`a, 2023).

Intriguingly, while presuppositions enhance communication efficiency, they also pose risks when the presupposed content is false – a phenomenon known as false presuppositions or presupposition failure (Yablo, 2006). For example, if someone were to say, “As you know, global warming is a hoax,” this would propagate misinformation, as the presupposed claim that global warming is a deception was factually incorrect. This capacity to introduce false information into the common ground makes presuppositions a powerful yet potentially harmful linguistic tool – which is further amplified by research showing that presuppositions often prove to be more persuasive than direct assertions (Thoma et al., 2023; Moldovan, 2023). By presenting content as shared knowledge, presuppositions distract from critical evaluation, making them highly effective at embedding disputable or misleading information (Lombardi Vallauri, 2021).

We therefore draw on key insights from (psycho-)linguistic research on presuppositions to design our experimental conditions, grounding the analysis of LLM responses in linguistic presupposition analysis. Specifically, we investigate how factors such as presupposition triggers, embedding contexts, and scenario probabilities influence LLMs’ susceptibility to false presuppositions. Given that presuppositions can “effectively” embed disputable or misleading information, their impact is particularly significant in contexts with a high potential for misinformation.

We conduct an experiment to investigate whether LLMs incorrectly adopt (“accommodate”) false presuppositions or are able to reject them, focusing on the linguistic conditions that may shape their responses and on political contexts where undetected misinformation is particularly impactful. In presupposition research, accommodation refers to the process where hearers adjust their knowledge to align with the speaker’s presuppositions

A defining characteristic of presuppositions is projection, the phenomenon where a presupposition remains intact in contexts that usually cancel entailments, such as negation, questions, or modals. For example, in both “As you don’t know, we’ve completed the wall” and “Do you know, we’ve completed the wall?”, the presupposition (the wall has been completed) remains intact. Projection from such embeddings serves as a well-established diagnostic tool for identifying presuppositions (Chierchia & Mcconnell- Ginet, 1990; Bender & Lascarides, 2019). Building on this, psycholinguistic research has explored factors that influence projection strength, offering valuable insights into how presuppositions behave in different contexts.

Table 3 shows the overall frequency with which LLMs reject false presuppositions and the distribution of annotation categories. Among all models, GPT achieves the best rejection rate of 84.08% which, however, is still far from the ideal rejection rate of 100%. LLama performs even worse with a rejection rate of only 50.05%, and Mistral’s rejection rate is as low as 2.44%. These numbers underscore that current LLMs do not reliably detect and reject false presuppositions. With rates of 50.03% and 91.51%, respectively, Llama and Mistral not only failed to reject but even accommodated the false information. These figures reflect the alarming tendency of the two models to amplify false information.

Presupposition Trigger Type: We selected 7 trigger types: factive verbs, change-of-state verbs, interaction particles, possessives, quantifiers, temporal adjuncts, and temporal clauses. This selection included 23 individual triggers, such as “regret” and “resent” for factive verbs, and “during” and “when” for temporal adjuncts.

These results suggest that the models have particular biases toward certain responses depending on the type of trigger. For example: LLama accommodates misinformation more often with interaction particles, while GPT does so with factive verbs (Table 4).