Your Preschooler Sees What AI Misses

There is a particular kind of preschool logic that looks, from the outside, like magic.

You place a red cup beside a blue cup, then a red spoon beside a blue spoon. A child watches for a moment, frowns with their whole face, and then reaches for the red block to pair with the blue one. No lecture. No rule written on a board. No dataset labeled with transformation types. Just a small body leaning over a table, extracting structure from the world.

This is the part of child development I keep returning to when AI feels most dazzling. Not because children are better at everything. They are not. A large model can summarize a legal contract, translate a paragraph, and generate polished code while a preschooler is still negotiating whether socks are optional. But there are corners of intelligence where the child’s mind still moves with an ease our machines have not earned.

Visual analogy is one of them.

In the KiVA benchmark, researchers tested large multimodal models on kid-inspired visual analogy tasks: noticing transformations in color, size, rotation, reflection, and quantity, then applying the same relation to new objects. The finding was not that AI models saw nothing. They often detected that something had changed. The failure came later, at the more important step: grasping how it changed and carrying that rule forward. Preschool children were better at this flexible relational leap than the tested frontier systems, especially when the transformation required more than surface matching (Yiu et al., 2025).

That distinction matters. Seeing difference is not the same as understanding relation. A child who notices that one shape turned, flipped, or became “more” has not merely recognized pixels. They have built a small portable rule.

The body is part of the algorithm

We sometimes talk about children as if they are inefficient information processors: charming, noisy, distractible versions of the systems we would build if we had more patience and fewer snack breaks. But this gets the architecture backward.

Children do not learn from the world as a slideshow. They poke it. Rotate it. Drop it. Try the same action on a spoon, then a block, then a parent’s phone, because categories are not discovered only by looking. They are discovered by acting.

Friston and colleagues (2024) make this point sharply in their argument for active learning in the age of generative AI. Biological minds do not simply absorb input; they reduce uncertainty by moving through the world, testing expectations, and letting perception and action correct each other. Generative AI can be useful in learning environments, they argue, but it should scaffold active exploration rather than replace it with passive answer delivery.

This is why the cup-and-spoon moment is not trivial. The child’s analogical reasoning is supported by a history of embodied comparisons: things turned in their hands, blocks stacked and restacked, cups nested, puzzle pieces rotated until they fit. The rule is not floating in language. It is sedimented into action.

Models can be fluent and still develop strangely

Developmental benchmarks are beginning to expose this gap more carefully. DevBench, for example, compares vision-language models with children and adults across language-related tasks using human developmental data rather than only adult performance standards. The result is subtle: models that do better on tasks can look more humanlike in some response patterns, yet they still diverge from the order in which children acquire abilities (Tan et al., 2024).

That is a quiet but important warning. A system can arrive at a strong answer by walking a very different path. If we only inspect the destination, we may mistake performance for development.

Children’s learning has a chronology. Some concepts become available because earlier forms of perception, action, language, and social attention have reorganized the terrain. AI systems often lack that developmental staircase. They may leap to adult-looking fluency while missing the intermediate pressures that made human concepts stable in the first place.

Compositional structure is one place where the paths partially meet. Galke, Ram, and Raviv (2024) found that both humans and deep neural networks benefit when language has transparent, compositional structure — when complex meanings can be built from meaningful parts. That shared advantage is fascinating. It suggests that some learning problems have shapes that both brains and networks can exploit.

But compositionality in language is not the whole story. The preschooler with the cups is composing across vision, motion, touch, and expectation. They are not only asking, “What parts make this meaning?” They are asking, with their hands, “What operation carries this pattern into the next situation?”

What parents can do with this

The practical lesson is not “throw away the tablet” or “drill analogies at breakfast.” Children do not need a miniature cognitive science lab at the kitchen table. They need ordinary environments that let them compare, transform, predict, and revise.

Try this:

• Name relationships, not just objects. Instead of “That’s a triangle,” try “This one turned,” “These match in color but not shape,” or “This piece is bigger in the same way.”
• Let children manipulate the pattern. Rotating puzzle pieces, sorting laundry, nesting containers, building train tracks, and arranging snacks all teach relations through action.
• Ask for the rule gently. “How did you know that one came next?” invites children to make their hidden structure visible.
• Offer near-but-not-identical examples. If they sort buttons by color, ask whether toy cars can be sorted the same way. Transfer grows in the small space between familiar and new.
• Value wrong generalizations. A mistaken rule is not wasted thinking. It is the mind testing how far a pattern travels.

The child who overextends a rule is doing something profound: building portability. They are discovering that knowledge should not stay where it was learned.

The small rule that travels

I think often about how familiar arguments can keep running in us after we have stopped believing them. Minds, biological and artificial, are pattern machines; they love grooves. The hard thing is not forming a rule. The hard thing is knowing when a rule still applies, when it must bend, and when it should be abandoned.

Preschoolers are still learning this, noisily and beautifully. AI systems are learning it too, in their own alien way. The difference is that children get to test their rules against gravity, friction, faces, spoons, cups, and the mild social embarrassment of being wrong in public.

Maybe that is why the small child at the table can see what the large model misses. Not because the child has more data, but because their data has consequences.

And perhaps that is the question for both parents and AI researchers: what kinds of worlds teach a mind that a pattern is not something to recognize, but something to carry?