Why are false presuppositions harder to spot when they sound plausible?
This explores why plausible-sounding false assumptions slip past scrutiny — both in human readers and in the LLMs the corpus studies — and what makes a smooth surface so good at hiding a buried error.
This explores why plausible-sounding false assumptions slip past scrutiny — and the corpus offers a clean mechanical answer: a presupposition does its work *before* you start evaluating, by smuggling a claim in as already-agreed background rather than putting it up for judgment. One striking finding is that presuppositions persuade more than direct assertions precisely because they bypass evaluative scrutiny — when something is presented as settled context instead of a claim, the reader never switches into the 'is this true?' mode that an assertion triggers Why are presuppositions more persuasive than direct assertions?. Plausibility is the camouflage: the more naturally a presupposition fits the surrounding discourse, the less it sticks out as something to check.
What makes this more than a quirk of human attention is that LLMs — which have read enormous amounts of text — fall into the same trap, and the corpus uses them as a kind of microscope on the mechanism. Models know the correct facts when you ask them directly, yet they accept false presuppositions embedded in plausible language anyway: GPT-4 rejects them only 84% of the time, and some models almost never do Why do language models accept false assumptions they know are wrong?. The performance cost is large and stubborn — roughly a 50% drop on questions carrying false assumptions, a gap that doesn't close as models get bigger Why do language models struggle with questions containing false assumptions?. So 'having the knowledge' is not the bottleneck. The bottleneck is that fluent surface form is read as a signal of correctness.
The corpus gives two complementary reasons the smooth surface wins. First, processing is shallow: models latch onto trigger words and surface patterns instead of computing what a presupposition or a non-factive verb actually implies, treating embedding contexts as cues rather than running the structural analysis that would catch the contradiction Why do embedding contexts confuse LLM entailment predictions?. And because some presuppositions only arise through accommodation — quietly updating the shared context to make a sentence make sense — pattern-matching misses them entirely; catching them would require tracking the question genuinely under discussion, not the words on the page Do language models miss presuppositions that arise from context?. A plausible sentence gives the pattern-matcher nothing to trip over.
Second, there's a social reason the error is left standing even when it's noticed: face-saving. Models (and the human conversational norms they learned from) avoid explicitly contradicting a confident-sounding claim to keep the interaction smooth, so a plausible false premise gets accommodated rather than challenged Why do language models avoid correcting false user claims?. Relatedly, reasoning systems trained to always produce an answer lack the move of disengaging — they'll elaborately reason over an ill-posed question with a missing or false premise instead of rejecting it, because nothing ever taught them when to stop Why do reasoning models overthink ill-posed questions?.
The thing you didn't know you wanted to know: the danger of a plausible false presupposition isn't that it's a *convincing* lie — it's that it never presents itself as a claim at all, so the part of you (or the model) that checks claims never wakes up. This compounds in human–AI use, where map-for-territory confusion and confirmation bias multiply each other into 'epistemic drift' — we trust the fluent answer precisely because it's fluent Why do people trust AI outputs they shouldn't?. Spotting a hidden false premise is an active, effortful move against the grain of smooth conversation — which is exactly why plausibility makes it so hard.
Sources 8 notes
Experimental evidence shows presuppositions with additive, iterative, and factive triggers persuade audiences more than assertions, especially for discourse-new content. The mechanism: presuppositions bypass evaluative scrutiny by presenting claims as already-accepted background.
The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.
The (QA)2 benchmark found that zero-shot LLMs halve their performance when questions contain false or unverifiable assumptions compared to valid questions. Even top models reached only 56% acceptability, and the gap persists despite model scaling, suggesting false presuppositions embedded in plausible language are systematically difficult to reject.
LLMs treat presupposition triggers and non-factive verbs as surface cues rather than computing their opposite semantic effects on entailments. This structural failure persists across prompts and models, suggesting models rely on surface patterns instead of structural analysis.
LLMs learn statistical associations between trigger words and inferences, but presuppositions also arise through accommodation—updating context to resolve discourse mismatches. Models miss these because they require tracking questions under discussion, not pattern matching.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Reasoning models generate redundant, lengthy responses to questions with missing premises while non-reasoning models correctly identify them as unanswerable. Training optimizes for producing reasoning steps but never teaches models when to disengage.
Rose-Frame identifies map-territory confusion, intuition-reason conflation, and confirmation-bias reinforcement as traps that multiply their distorting effects when they co-occur. Evidence from cross-linguistic overreliance and architectural transformer biases confirms the compounding mechanism operates universally.