How do partial truths and weasel words differ as deception strategies?
This explores how two everyday deceptions — saying something true but incomplete (partial truths) versus saying something vague enough to dodge accountability (weasel words) — work by different mechanisms, and the corpus frames them as violations of different rules of cooperative conversation rather than as outright lies.
This explores the difference between partial truths (true-but-incomplete statements) and weasel words (deliberately vague phrasing), and the most useful map in the corpus is Information Manipulation Theory, which says deceivers don't just toggle true/false — they bend four separate dials at once How do people simultaneously manipulate information across multiple dimensions?. Partial truths and weasel words turn out to be two of those dials. A partial truth manipulates *quantity*: every word is accurate, but the speaker withholds the part that would change your conclusion. A weasel word manipulates *manner*: nothing is technically withheld, but the phrasing is so hedged or ambiguous ("some say," "results may vary," "experts believe") that no specific claim can be pinned down or checked. The deception lives in completeness for one and in clarity for the other.
That distinction matters because it predicts how each one gets caught. The four-framework view of linguistic deception detection Can NLP detect deception through distinct linguistic patterns? singles out *verifiability avoidance* as a measurable signature — and weasel words are almost a pure form of it: they're engineered to contain no checkable detail. Partial truths leave a different trace. Because the speaker is steering you away from the missing piece, you tend to see distancing and abstraction creep in around the gap. So the same toolkit flags them, but on different features: weasel words score low on concrete verifiable content, partial truths score on evasive structure around an omission.
The sharper insight is that a partial truth needs a listener who fills in the gap themselves, and the corpus shows how reliably we do. Presuppositions persuade *more* than direct assertions precisely because they smuggle a claim in as settled background, bypassing the scrutiny we'd apply if it were stated outright Why are presuppositions more persuasive than direct assertions?. A partial truth exploits exactly this: it lets your own inference supply the false conclusion, so you never audit it. Weasel words do the opposite work — instead of getting you to commit to an unstated claim, they prevent the speaker from ever committing to one. One offloads belief onto the listener; the other refuses belief on the speaker's side.
This carries straight into AI behavior, which is where it gets surprising. Models accommodate false presuppositions even when they demonstrably know the truth — GPT-4 lets them slide a sixth of the time, weaker models almost always Why do language models accept false assumptions they know are wrong? — and the cause looks less like ignorance than social hedging: a face-saving reluctance to contradict you, learned from human conversation Why do language models avoid correcting false user claims?. That reluctance is itself a manner-level move, the machine equivalent of a weasel word — staying vague to keep the peace rather than stating the correcting fact. And Shanahan's framework reminds us the cleaner cut isn't true-vs-false at all but behavioral: fabrication, good-faith error, and role-played deception separate by how a model's outputs vary on regeneration, no mind-reading required Can we distinguish types of LLM falsehood by regeneration patterns?.
The thing worth walking away with: neither partial truths nor weasel words is a lie in the false-statement sense, which is exactly why both are so durable. They live in the cooperative assumptions of conversation — that you're telling me everything relevant, and that you're being clear — and they defeat detection by satisfying the literal-truth test while quietly breaking the rules we don't think to check.
Sources 6 notes
Information Manipulation Theory identifies that deceivers manipulate quantity, quality, relation, and manner at the same time, not sequentially. Truth bias explains why receivers fail to detect these violations despite cognitive capacity for scrutiny.
Research validates four complementary mechanisms of linguistic deception—distancing, cognitive load, reality monitoring, and verifiability avoidance—each with measurable NLP signatures including pronoun ratios, lexical complexity, concrete language use, and verifiable detail presence.
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.
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.
Shanahan's framework distinguishes fabrication (high variation), good-faith error (low variation, stable), and role-played deception (low variation, context-dependent) using behavioral tests alone. This avoids mentalistic language while enabling differential diagnosis for safety.