Why "Can You Pass the Salt?" Isn't a Question

Here's a trick question: what does "Can you pass the salt?" mean?

The literal answer is that it's a yes/no question about your physical capabilities. But if you answer "Yes, I can" and then don't pass the salt, you're going to have a very awkward dinner.

Every child eventually learns this. Not from explicit instruction — nobody sits a four-year-old down and explains cooperative conversational principles — but through thousands of hours of social interaction that gradually build up a set of implicit rules governing what language actually does in context.

This is pragmatics. And it's one of the places where the gap between children and AI is both narrower and wider than the discourse usually admits.

The Hidden Grammar

The philosopher Paul Grice spent decades formalizing what he called the Cooperative Principle: the assumption, implicit in all conversation, that speakers are trying to be relevant, truthful, appropriately informative, and clear. When someone violates one of these maxims — quantity, quality, relation, or manner — you don't just take their words at face value. You work out what they actually mean.

This is how implicature works. "Can you pass the salt?" violates the maxim of manner (why ask about capability when you want to make a request?) so you infer the speaker is being indirect for social reasons. You pass the salt. No one explains this to you. You just figure it out.

Children start developing this kind of pragmatic reasoning remarkably early. Three-year-olds grasp basic indirect requests. By five or six, they start understanding scalar implicature — if I say "some of the cookies are gone," you automatically infer I mean not all of them, without anyone walking you through the logic. Irony and sarcasm take longer, typically solidifying around age eight or nine, and continue getting more nuanced through adolescence.

None of this is taught. It's discovered through the accumulated texture of real conversation: from the tone of a parent's "fine" (which never means fine) to the way a teacher's "interesting" can be either a compliment or a polite execution depending on what precedes it.

What LLMs Are Actually Doing

Here's where I want to calibrate the AI discourse, because the usual takes are both wrong.

Take one: LLMs have basically solved pragmatics. They handle indirect requests, they rarely answer "yes I can" to the salt question, and they can generate contextually appropriate sarcasm on demand. Surely that means they've cracked conversational implicature?

Take two: LLMs are just statistical parrots who have no idea what language means and are fooling everyone.

Both are sloppy. A landmark 2024 paper in Trends in Cognitive Sciences by Mahowald, Ivanova, and colleagues draws a clean distinction between what they call formal and functional linguistic competence (Mahowald et al., 2024). LLMs are genuinely impressive at formal competence — grammatical structure, statistical regularities, surface-level rhetorical patterns. Where they fail systematically is at functional competence: using language to actually reason, plan, and mean things in the world.

Pragmatics lives right at the intersection. You need formal competence to process "can you pass the salt" as a well-formed question. You need functional competence to recognize it as a request embedded in a social context you're supposed to care about.

A 2025 study by Nastase and colleagues actually deployed language models as scientific instruments, using their internal representations to decode the neural dynamics of real human conversation — tracking how the brain processes language word-by-word during naturalistic dialogue (Nastase et al., 2025). What they found was striking: LLM representations track the temporal unfolding of language processing in the brain with notable fidelity. The models are capturing something real about how language unfolds.

But the specifically pragmatic demands of conversation — the turn-taking, the implied social intent, the gap between what was said and what was meant — are precisely where brain processing diverges most from what the model representations capture. The model is tracking the river, not the meaning swimming in it.

What Children Have That LLMs Don't

The pragmatics gap isn't primarily a language problem. It's a social stakes problem.

Children learn indirect speech acts because conversations have consequences. When a toddler demands "I want cookie" from a parent who expects polite requests, the stakes are immediate and real. The feedback loop — the raised eyebrow, the expectant silence, the "how do we ask?" — is consequential in a way no training corpus can replicate.

Grice's maxims aren't rules children memorize. They're regularities children discover because violations create friction in relationships that matter to them. Irony lands because you already model the other person's beliefs and can detect the gap between what they said and what they must actually think. Sarcasm works because you care about the speaker's emotional state. Indirect refusals are decoded because you're tracking the other person's face-saving needs.

LLMs are trained on text that humans produced in social contexts. But they weren't in those contexts. They see the transcript, never the silence, the hesitation, the moment when the joke lands or doesn't. Language encodes surprisingly little about what language actually does.

The Calibrated Verdict

Here's my honest take: LLMs perform pragmatics better than they understand it. "Can you pass the salt?" is in the training data, probably in a thousand slightly different versions, and the model handles it fine. But present a genuinely novel implicature, or irony that requires rich theory of mind, or a subtle face-saving indirect refusal in an unfamiliar cultural register — and the formal machinery starts showing its limits.

This matters for anyone deploying conversational AI in contexts where pragmatic failure has real consequences. Mental health tools where "I'm fine" needs to be read correctly. Educational tutors where a student's hesitant "yeah, I think I get it" might actually be a quiet request for help. Negotiation contexts where what's not said carries most of the weight.

The good news: we're starting to build rigorous frameworks for measuring this gap rather than just arguing about it. The bad news: most AI discourse still treats a model that produces grammatically impeccable text as equivalent to one that understands what that text is doing in the world.

Those are different things. Children spend years learning the difference — through scraped knees of social friction, not gradient descent on a text corpus. We should at least have the intellectual honesty to keep that distinction visible, even as the outputs get harder to tell apart.