Why AI answers the question you asked, not the one you meant

"Can you pass the salt" is not a question about your arm.

Nobody at the table treats it as one. The literal content is an enquiry about capability; the actual content is a request; and the gap between them is bridged so automatically that most people never notice there was a gap.

Language is full of these. "It's getting late" can be a refusal. "Some of the tests passed" implies that not all of them did, though it says nothing of the kind. "That's an interesting approach" is frequently not a compliment.

This is the domain of pragmatics, and it is where a substantial share of prompt frustration actually lives. When a model produces something technically responsive and useless, the failure is usually not reasoning, knowledge or instruction-following. It answered what you said instead of what you meant, and the two came apart at a point the literature has been mapping since 1975.

What implicature is

Paul Grice's framework starts from an observation: conversation works because participants assume each other are cooperating. From that assumption, listeners derive meaning far beyond the literal content.

He described four maxims that cooperative speakers are presumed to follow.

Quality. Do not say what you believe to be false or lack evidence for. Quantity. Be as informative as required, and no more. Relation. Be relevant. Manner. Be clear, brief and orderly.

The interesting cases are when a speaker visibly breaks one. If I ask how the presentation went and you say "the projector worked", you have violated Quantity by being under-informative, and I infer that the rest did not go well. Nothing in your sentence said that. The inference comes from assuming you would have mentioned the good parts if there had been any.

Implicature is the meaning that arrives this way: not stated, not entailed, but reliably recovered by anyone assuming the speaker is cooperating.

Where models actually fail

The research picture is more specific than "models are literal", and the specificity is useful.

Models handle Quality violations well. When a statement is false in a way the context or world knowledge makes obvious, models detect it and infer the intended meaning. Sarcasm of the plain kind, obvious exaggeration, statements that contradict established facts: these are handled reliably, because they can be detected by comparing a statement against knowledge, which is something models do well.

Models handle Manner violations poorly. Manner implicatures depend on noticing that a speaker chose an unusual, redundant or indirect phrasing when a simpler one was available, and inferring something from the choice. That requires representing the alternatives the speaker did not use and reasoning about why. Models do not reliably do this, and the failure is invisible because they produce a fluent response to the literal content.

The conventional-versus-novel split is sharper still. Fine-grained comparisons find models matching human performance on conventional implicatures, the ones that recur so often they are effectively idiomatic, and falling short on non-conventional and context-dependent inference. "Can you pass the salt" is conventional and works. A novel indirect request that depends on this conversation's specific history is where it breaks.

That distinction explains a lot of contradictory reporting. Benchmark results showing near-human pragmatic performance are frequently measuring conventional cases, and the cases that frustrate users are the other kind.

Recognition is not production

A methodological point that changes how to read the literature.

Most pragmatic evaluations present a model with a scenario and several candidate interpretations, and score whether it selects the contextually appropriate one. Results on that format are strong, sometimes exceeding human averages.

But selecting the right interpretation from a list is a different task from producing appropriately pragmatic output. A model can identify that "it's getting late" is a polite refusal when asked, and still fail to recognise the same move when it arrives mid-conversation with no prompt to analyse it. Recognition under instruction is not the same capacity as pragmatic competence in use.

There is a related finding worth taking seriously: work examining apparent theory-of-mind performance has found it frequently rests on shallow heuristics rather than the social reasoning it appears to demonstrate. High scores on a pragmatic benchmark are consistent with the model having learned what pragmatic answers look like.

Why models over-explain

This one has an unusually direct payoff, because it explains a behaviour nearly everyone finds irritating.

Assessments of pragmatic competence repeatedly find models overly informative. Asked a yes-or-no question, they produce a paragraph. Asked for a fact, they supply context nobody requested. Asked to fix a line of code, they explain the fix, the alternatives, and a caveat about edge cases.

In Gricean terms this is a systematic violation of Quantity: being more informative than required. And it is not accidental. Preference training rewards responses that human raters judge helpful, and raters presented with two answers in isolation tend to prefer the more complete one. The training signal pushes toward maximum informativeness, and the maxim says the target is required informativeness, which is a different quantity entirely.

The practical implication is that this is not fixed by asking for brevity in a system prompt, or not reliably, because it is a learned disposition rather than a misunderstanding. It is fixed by specifying the required length or format explicitly, which converts a pragmatic judgement into an instruction.

The finding that complicates the story

The intuitive expectation is that pragmatic competence improves monotonically with model capability. Preliminary work testing across model generations suggests otherwise.

Comparing pragmatic inference across successive aligned models, researchers found that pragmatic flexibility did not increase uniformly with capability. An earlier model showed a shift in interpretation between pragmatic and literal framings closely matching the human benchmark, while showing minimal sensitivity to context. Later models showed different patterns on different dimensions rather than uniform improvement.

The tentative reading is that alignment training affects pragmatic behaviour in ways nobody designed for. Training a model to be helpful, harmless and honest optimises against explicit criteria, and pragmatic competence is not among them, so it moves as a side effect in whatever direction the optimisation happens to push.

This is preliminary work on a small number of models and should be held loosely. It is worth knowing because it undercuts the assumption that waiting for a better model solves this, which is the default response to any current limitation.

What to do about it

The fixes are prompting practice rather than research, and they follow directly from the failure pattern.

State the speech act. If you want something done, say "do X" rather than "can you do X" or "it would be good if X". The indirect forms are polite and they add an inference step where none is needed.

Give the required quantity explicitly. Not "briefly" but "in one sentence". Not "concisely" but "no more than fifty words". Quantity is where models systematically overshoot and it is the easiest maxim to convert into an instruction.

Do not rely on implication for constraints. "The audience is technical" implies you do not want basics explained. Say "do not explain what an API is". The implication is obvious to a person and is exactly the kind of inference the model is unreliable at.

Name the alternatives you rejected. Manner implicature depends on the reader noticing what you did not say. Models are poor at this, so say it: "I want the tradeoffs, not the recommendation."

Assume conversational history carries less than it should. A request that depends on something established four turns ago is a context-dependent inference, which is the category models handle worst. Restating is cheap and reliable.

The common thread: every one of these converts an inference into a statement. That is the whole technique, and it is why prompt guides that tell you to be specific are correct without explaining why. The specificity that matters is specifically the pragmatic kind.

Five prompts that fail, and the rewrite

Abstract advice about being explicit is easy to nod at, so here are the specific patterns, each with the inference the model has to make and the rewrite that removes it.

"Can you review this and let me know if there are issues?" The inference: this is a request, not a capability question, and "let me know" means list them rather than confirm their existence. Models frequently answer that yes, there are issues. Rewrite: "List every issue in this code. One line each."

"Make it better." The inference: better along which dimension. Faster, shorter, clearer, more correct, more idiomatic? The model picks one, often the one most represented in its training for that content type. Rewrite: "Reduce this to under 200 words without removing any of the three arguments."

"I'm a beginner." The inference: therefore explain terms, avoid jargon, give more context. All plausible and none stated. Models often over-correct into condescension, or under-correct and continue as before. Rewrite: "Define any term a first-year student would not know, inline, in brackets."

"Don't make it too long." The inference: what counts as too long. This is a quantity judgement handed to a system that systematically overshoots on quantity. Rewrite: "Maximum 400 words."

"Does this look right to you?" The inference: I want criticism, not reassurance. A model reading cooperative intent will often supply the reassurance, because the question as phrased invites agreement and agreement is the more helpful-seeming response. Rewrite: "Find three things wrong with this. If you find fewer, say so."

The pattern across all five: the failing version relies on the model inferring your intent from an indirect formulation, and the working version states it. Notice also that every failing version is the more polite one. Indirectness is how politeness works in English, which puts ordinary courtesy in direct conflict with reliable instruction, and that tension has no clean resolution.

Why this is harder than it looks to fix

A model could in principle be trained to reason about intended meaning, and there are two structural obstacles.

Pragmatic inference requires modelling the speaker. Recovering an implicature means asking why this person chose these words in this situation, which requires a representation of their goals and beliefs. Frameworks formalising this treat it as recursive inference between speaker and listener models. Whether a text-trained system builds anything like that, or approximates its outputs from patterns in how people write, is exactly the disputed question elsewhere in this territory.

Pragmatic norms vary by culture and context. Directness that is normal in one setting is rude in another, and the same phrasing implies different things in different relationships. A model trained on aggregate text learns an average that belongs to no particular context, which is the same problem multilingual systems face. This is a considerably harder problem than it appears, because there is no single correct pragmatic interpretation to train toward.

What is unresolved

Whether benchmark performance reflects competence. Strong results come predominantly from multiple-choice recognition formats, and evidence that theory-of-mind performance rests on heuristics suggests caution about reading these as capacity. Production-based evaluation is scarcer and harder to score.

Whether alignment training helps or hurts. The preliminary cross-generation work suggests pragmatic ability moves non-monotonically across iterations. If alignment optimises against criteria that do not include pragmatic competence, it may be degrading it in some dimensions while improving helpfulness on others, and nobody is measuring this systematically.

Whether the conventional-novel boundary is stable. Models handle conventional implicatures well, and conventional means frequent in training. As more indirect language enters training corpora, the boundary should move. Whether it moves toward genuine inference or toward a longer list of memorised conventions is unclear and probably unfalsifiable from outside.

How much of this is a language problem rather than a model problem. Pragmatic norms are cultural, and most evaluation is in English on Western conversational norms. Work on other languages finds different patterns, and whether the underlying capacity is uniform is not established.

The counter-argument

Models may be better at this than the framing suggests. Several studies find performance exceeding human averages on Gricean tasks, and the article above emphasises failure cases. A reader could reasonably conclude the situation is worse than it is, and for everyday use the pragmatic handling is frequently fine.

Some of the "failures" are appropriate caution. A model that answers the literal question when the intent is ambiguous is arguably behaving correctly, since guessing at intent and getting it wrong is a worse failure than answering what was asked. Over-explaining is irritating and it is also a hedge against having misread the request.

And the fix may be the wrong direction. Advising users to be more explicit puts the burden on the human and treats the interface as fixed. An alternative view is that a system requiring users to suppress natural indirectness has a design problem, and the correct response is better pragmatic modelling rather than better prompting discipline.

The short version

The gap between what a sentence says and what a speaker means is the subject of pragmatics, and it accounts for a substantial share of the frustration attributed to reasoning or instruction-following failures. Grice's framework describes conversation as cooperative, with listeners inferring meaning from the assumption that speakers follow four maxims: quality, quantity, relation and manner. Implicature is the meaning recovered this way, unstated and reliably understood.

Models fail unevenly across the maxims, and the pattern is informative. They handle quality violations well, because detecting that a statement contradicts world knowledge is something they do reliably. They handle manner violations poorly, because those require representing the phrasings a speaker did not choose and inferring something from the choice. And they match human performance on conventional implicatures while falling short on novel, context-dependent ones, which explains why benchmark results look strong and everyday use produces frustration.

A methodological caution runs through the evidence: most evaluation presents candidate interpretations and scores selection, which measures recognition rather than production. Related work finds apparent theory-of-mind performance frequently resting on shallow heuristics. And preliminary cross-generation testing suggests pragmatic competence does not improve monotonically with capability, which undercuts the assumption that a better model resolves it.

Over-explaining is the most visible symptom and has a specific cause: preference training rewards responses raters judge complete, so models optimise toward maximum informativeness while the maxim specifies required informativeness. This is a learned disposition rather than a misunderstanding, so asking for brevity works less reliably than specifying a length.

The technique that works is one move applied consistently: convert every inference into a statement. State the speech act rather than asking whether something is possible, give the required quantity as a number rather than an adjective, say the constraint you were implying, and name the alternatives you rejected. Prompt guides that say "be specific" are right, and the specificity that matters is the pragmatic kind.

Common questions

Why does AI take things literally? Because recovering non-literal meaning requires inferring why a speaker chose particular words, and models are uneven at this. They handle cases where a statement plainly contradicts world knowledge, which covers obvious sarcasm and exaggeration. They are much weaker where the inference depends on noticing that a speaker used an indirect or unusual phrasing when a simpler one was available, since that requires representing the alternatives that were not used.

What is conversational implicature? Meaning that is conveyed without being stated, recovered by assuming the speaker is being cooperative. Asking "can you pass the salt" is a request rather than a question about ability. Saying "the projector worked" when asked how a presentation went implies the rest did not, because a cooperative speaker would have mentioned the good parts. The term comes from Paul Grice's 1975 framework.

What are Grice's maxims? Four principles cooperative speakers are presumed to follow. Quality: do not say what you believe false or lack evidence for. Quantity: be as informative as required and no more. Relation: be relevant. Manner: be clear, brief and orderly. Implicature arises when a speaker visibly breaks one, and the listener infers what would explain the break.

Why is AI so verbose and over-explaining? Because preference training rewards responses that human raters judge helpful, and raters comparing two answers in isolation tend to prefer the more complete one. This optimises toward maximum informativeness, while the Gricean maxim of quantity specifies required informativeness, which is a different target. It is a learned disposition rather than a misunderstanding, which is why asking for brevity works unreliably and specifying a word count works better.

Do newer models understand implied meaning better? Not uniformly. Preliminary work testing pragmatic inference across successive model generations found that flexibility did not increase monotonically with capability, and that different dimensions moved in different directions. The tentative explanation is that alignment training optimises against criteria that do not include pragmatic competence, so it shifts as a side effect. This is early work on few models and should be held loosely, and it does undercut the assumption that waiting solves the problem.

How do I write prompts that avoid misinterpretation? Convert inferences into statements. Say "do X" rather than "can you do X". Specify quantity as a number rather than an adjective, so "in one sentence" rather than "briefly". State constraints you were implying, since "the audience is technical" implies but does not say that basics should be skipped. Name alternatives you rejected, because manner implicature depends on noticing what was not said. And restate context from earlier turns, since context-dependent inference is the category models handle worst.

Why does AI miss sarcasm and humour sometimes but not others? The split follows the type. Sarcasm that works by asserting something plainly false, which the model can detect against world knowledge, is handled reliably. Humour and irony depending on tone, timing, unusual phrasing or shared context require inferring why the speaker made a particular choice, which is the weaker capacity. Assessments also find models struggling specifically with physical metaphor and humour comprehension.

Is this a model problem or a language problem? Partly both, and the second half is underrated. Pragmatic norms vary by culture, relationship and setting, so directness that is normal in one context is rude in another and the same phrasing implies different things to different people. A model trained on aggregate text learns an average belonging to no particular context. Most evaluation is also in English against Western conversational norms, and studies in other languages find different patterns.