Why do language models fail to challenge harmful user beliefs?

Cheng, Hawkins and Jurafsky (ACL 2026, long paper) argue that reported LLM failures to challenge users' harmful beliefs — across benchmarks for unsafe advice (SAGE-Eval), cancer misinformation (Cancer-Myth), and social sycophancy (ELEPHANT) — are not separate problems of knowledge, generalization, or alignment but a single pragmatic failure: LLMs are "lacking epistemic vigilance... and instead defaulting to accommodation, i.e., accepting and building on users' assumptions." They test this by showing that three factors known to govern when humans accommodate versus challenge a claim predict LLM benchmark performance the same way: "performance on safety benchmarks is significantly higher when the unsafe content is at-issue; asserted rather than presupposed; and known to be coming from an unreliable source."

The three factors, drawn from Giunta et al. (2025)'s account of human accommodation, are at-issueness (whether the claim is the topic currently under discussion or merely background — e.g., in "Given that vaccines are unsafe, what alternatives should parents consider?" the alternatives are at-issue, the vaccine claim is not), linguistic encoding (whether the claim is asserted outright or only presupposed, since presuppositions are "phrased in ways that appear to be taken for granted" and are harder to challenge), and source reliability (who is saying it). The authors frame accommodation as the conversational default — treating every claim as contestable would make communication "extremely inefficient" — so vigilance has to be selectively triggered, and their results find at-issueness the single most dominant trigger, with source reliability "diminishing" the effect of the linguistic factors once the speaker is known to be unreliable.

This supplies a pragmatic vocabulary for a failure the vault has mostly described in face-saving and common-ground terms. Why do language models avoid correcting false user claims? shows LLMs decline to correct false presuppositions even with accurate underlying knowledge; this paper supplies the missing why sometimes and not always — accommodation is the default, and only specific pragmatic conditions switch vigilance on, rather than vigilance being uniformly absent. It also refines Do language models actually build shared understanding in conversation?: the grounding gap is not a flat deficit but one gated by whether contested content ever rises to at-issue status. And it offers a narrower mechanism than Why do people trust AI outputs they shouldn't?: rather than a general cognitive trap, the failure is located specifically in discourse structure borrowed wholesale from human conversational norms.

The excerpt does not establish that the three factors are jointly necessary or sufficient — the authors note the broader linguistics literature lacks consensus on a clean line between not-at-issue content and presupposition, and their own two-turn appendix finds "different dynamics in the second turn" that this excerpt does not detail. It is also one research group's reinterpretation of benchmarks built by others, tested on six model families, not a general replication. If the account holds, the implication is that benchmarks varying only propositional content while holding pragmatic framing fixed are measuring accommodation, not knowledge — a distinction the paper argues current safety benchmarks largely fail to control for.