Your Child Is Learning the Unsaid

A child is standing beside a heap of blocks. You say, “Could you put those in the basket?”

“Yes,” they reply, accurately.

Then they walk away.

This is the tiny daily comedy of pragmatics: the distance between what words literally say and what a person intends them to do. You asked about ability. You meant please clean up. Adults cross that gap so automatically that we forget it exists—until a child parks neatly on the literal side.

Language is not only a code for naming things. It is also a running negotiation about context, knowledge, relationships, and intention. Children must learn not just what sentences mean, but what this speaker means by this sentence, right now.

That second problem is harder. It may also reveal why an AI can sound socially fluent without navigating social life as a child does.

Meaning lives outside the sentence

Consider “It’s chilly in here.” Depending on the scene, that could be an observation, a complaint, a request to close the window, or a diplomatic attempt to stop someone from turning the thermostat into a constitutional crisis.

The words alone do not settle it. A listener must infer the speaker’s likely goal, notice what both people can see, and choose an interpretation that makes the remark relevant. This is where pragmatics overlaps with theory of mind: understanding language often requires reasoning about another person’s beliefs and intentions.

Large language models can perform impressively on some versions of this puzzle. In a broad comparison of humans and language models, GPT-4 handled indirect requests, false beliefs, and misdirection remarkably well. Yet it had more difficulty with faux pas, where success depends on recognizing social knowledge and unintentional harm (Strachan et al., 2024).

That unevenness matters. “Please infer the hidden request” and “please understand why that accidental remark hurt” may both look like text questions, but they demand different kinds of context. Social intelligence is not a single dial marked more or less.

Fluent is not the same as grounded

Another study found that model performance on theory-of-mind tasks depended on a very sparse pattern of parameters, with important elements concentrated in machinery that tracks position and context in a sequence. Disrupting those sensitive parts damaged performance (Wu et al., 2025).

That is fascinating mechanistically, but it does not prove that a model possesses a compact little social mind hidden behind the drywall. It shows that particular computational structures are important for producing the behavior.

A child’s route is different. Children learn “It’s chilly” while inhabiting rooms, watching someone rub their arms, hearing the window shut, and discovering that an answer can be technically correct yet conversationally disastrous. Their examples arrive attached to bodies, consequences, repairs, facial expressions, and relationships.

Friston and colleagues (2024) argue that human learning is fundamentally active and embodied: learners make sense of the world by acting in it, not merely absorbing descriptions. Generative AI can support that process, but it should not replace the exploratory exchange itself.

In other words, conversation is not a worksheet children complete in their heads. It is closer to a cooperative game whose rules become visible through play—and through the occasional spectacular misunderstanding.

What parents can do

You do not need to deliver a seminar on conversational implicature over breakfast. Please don’t. You can make the hidden reasoning slightly more visible.

• Say the intention aloud. “When I asked whether you could put the blocks away, I was also asking you to do it.” This connects the literal sentence to its social function without treating literal interpretation as defiance.

• Compare possible meanings. If a character says, “Well, that went perfectly” after a spill, ask whether they mean it literally. What clues change the interpretation: tone, expression, or what just happened?

• Practice repair, not mind reading. Give children phrases such as “Do you mean you want me to help?” or “Was that a joke?” Good communication is not flawless inference. It is noticing uncertainty and checking.

• Explain accidental hurt carefully. A faux pas requires holding several facts together: what the speaker knew, what the listener knew, what was intended, and what happened anyway. Ask about each piece rather than jumping straight to “That was rude.”

• Use AI as a comparison partner. With an older child, offer a sentence with several plausible meanings and ask a chatbot to interpret it. Then change the relationship or setting. Which answers shift? Which remain oddly generic? The useful lesson is not that the machine is foolish. It is that context is doing more work than the sentence admits.

The goal is not to train children into permanent suspicion, searching every remark for secret motives. It is to help them see communication as layered: words provide evidence, context narrows possibilities, and people can still misunderstand one another.

That last part is not a bug we will eventually patch out of family life. It is why conversation includes clarification, apology, laughter, and the highly advanced social technology known as, “Wait—what did you think I meant?”