Kids Learn Together. AI Just Plays Alone.

I was at an elementary school maker morning a couple of weeks ago, nursing a coffee and watching a five-year-old try to get her motorized cardboard car to drive straight. Twenty minutes of solo trial-and-error: nudging wheels, frowning, nudging again. She was deep in her own proprioceptive feedback loop and I was already scribbling notes on a napkin.

Then a second kid walked over.

He didn't say anything explanatory. He just crouched down, poked the left rear wheel, watched the axle wobble, and said "I think that one needs to go forward a little." She slid it forward. The car drove straight. They high-fived. The whole thing took maybe forty-five seconds.

I flipped the napkin over and wrote: peer learning = shared sensorimotor reference frame. Because that's exactly what happened. Two kids, one problem, a wordless transfer of physical understanding that no adult instruction had achieved in twenty minutes of patient solo debugging.

That moment crystallized something I keep coming back to: learning in groups isn't just learning faster. It's a qualitatively different kind of learning, built on cognitive machinery that kids are born with and that AI systems don't have anything close to. Yet.

The Hidden Engine: Joint Attention

Before a child can learn from another person, they first have to learn with them — to share attention on the same thing at the same time. This is joint attention, and it's so fundamental that developmental scientists treat it as a cornerstone of social cognition.

A systematic review by Grossmann et al. (2025), synthesizing 16 neuroimaging studies with EEG, fNIRS, and fMRI across infants aged 8 to 24 months, pinpointed the right temporoparietal junction (TPJ) as the core brain region involved across all forms of partnered social interaction. The TPJ is doing triple duty here: mentalizing (modeling what the other person knows), perspective-taking (seeing what they're seeing), and attention regulation (tracking where their focus is going).

Think about what that means for a moment. When that second kid at maker morning crouched down to look at the car, both children were running something like a real-time simulation of each other's attention and intention. The first kid didn't just receive information — she received it from a shared vantage point, calibrated to her specific situation. That's not information transfer. That's co-construction.

Joint attention is exactly the shared-reference capability that social robots and conversational AI systems are still struggling to replicate, which tells you something about how non-trivial this biological solution actually is.

Your Baby's Brain Is Wired for Social Learning

Here's what makes this even more striking: the social wiring runs deeper than behavior. It goes all the way down to what the infant brain does when another person is simply present with them.

A landmark study by Bosseler et al. (2024) at the University of Washington's I-LABS used magnetoencephalography (MEG) to measure brain responses in 5-month-old infants — 5 months old! — during live face-to-face social interaction versus a non-social control condition. The finding: neural theta activity in right-hemisphere attention and sensorimotor regions, recorded during that live social interaction, predicted language development at five follow-up timepoints stretching from 18 all the way to 30 months. Over two years later.

Let that sink in. How well a 5-month-old's brain lit up during face-to-face interaction predicted how well they were learning to talk at age 2.5.

The researchers describe what's happening during that interaction as a "social ensemble" — infant-directed speech, contingent responses, eye contact, the whole feedback-rich package. What it does, neurobiologically, is prime the infant's attentional and sensorimotor systems for learning. Social interaction doesn't just occasion learning. It opens the brain up for it.

This is a big deal for anyone thinking about peer learning. It's not that children happen to learn well from other children because they speak the same developmental language (though that's true). It's that the brain has a dedicated mechanism for using social co-presence as a learning amplifier. The presence of an engaged other person — someone whose attention you can read, whose gaze you can follow, who responds contingently to what you do — activates a mode of cognition that solo experience simply doesn't.

What Multi-Agent AI Actually Does

Now here's where AI enters the picture, and I want to be precise about what it does and doesn't do — because there's a version of this story that goes "multi-agent AI learns collaboratively, just like children!" and that version is flattering but misleading.

The most impressive results in multi-agent AI research are genuinely remarkable. In a 2025 Nature paper, Oh et al. from Google DeepMind showed that populations of AI agents, iterating through meta-learning across complex environments, can discover their own reinforcement learning update rules — producing DiscoRL, a novel algorithm that outperforms all manually designed RL rules while using roughly 40% less compute. That's a system of agents that collectively stumbled upon something no individual human researcher had engineered. Wild.

Similarly, Koster et al. (2025) trained an RL agent to act as a social planner in a multiplayer trust game involving nearly five thousand real human participants. The agent discovered a counterintuitive but highly effective cooperation strategy — be generous when resources are abundant, sanction free-riders when they're scarce — that outperformed classical game-theoretic mechanisms for sustaining cooperation. Again, a multi-agent system discovering something surprising through collective interaction.

These are not small achievements. But notice what's happening in both cases: agents are playing against each other, or playing alongside simulated humans, in service of a reward signal. The population dynamic drives exploration and discovery. What's absent is something more fundamental — the kind of shared intentionality and mutual attentional scaffolding that's happening when two kids work on a cardboard car together.

The Gap That Actually Matters

Self-play and multi-agent RL generate diversity through competition and iteration. What they don't generate is joint attention — the shared referential frame that allows one learner to understand not just what another agent did, but what that agent was noticing, what they were trying, and what they understood.

When my neighbor's kid walked up to that cardboard car, he wasn't just observing the outcome of the other kid's actions. He was modeling her attentional focus, reading her frustration, identifying the specific physical feature she'd missed, and pointing her attention toward it. He was performing a kind of collaborative cognitive surgery in real time, using social signals — gaze, gesture, tone — as instruments.

Multi-agent AI systems share a reward structure. Children share a world. They share attention, intention, and a web of social expectations about what counts as helping versus showing off, what to imitate versus what to improve on, when to speak and when to just crouch down and poke the wheel. That infrastructure runs on the same TPJ-driven joint-attention circuitry that Grossmann et al. (2025) found developing across the first two years of life — and there's no clear analogue for it in current multi-agent architectures.

The DeepMind cooperation work is evidence that RL agents can converge on socially beneficial strategies when incentives are structured correctly. That's genuinely useful and important. But it's a different problem than the one infant peer learners are solving. They're not just finding equilibria. They're building a shared model of what the other person sees, knows, and intends — and then using that model to scaffold each other's cognition in real time.

What This Means for Parents (and Designers)

If you take the developmental science seriously, the practical implication is almost embarrassingly simple: protect time for peer learning. Not instruction. Not adult-guided play. Peer interaction — kids working through things together, with all the negotiation and confusion and "wait, why did you do that?" that comes with it.

The Bosseler et al. (2024) findings suggest that even very young children's brains are doing something fundamentally different when engaged socially versus working alone or receiving one-directional information. That doesn't mean solo learning is worthless. It means peer learning activates cognitive systems that other modes of learning don't reach in the same way. If your kid spends all their learning time in front of screens — even excellent, interactive, AI-driven screens — they're probably leaving some of that social-learning machinery underactivated.

And if you're thinking about AI tutors or collaborative learning tools for kids: the bar isn't "does the AI give good feedback?" It's "does the interaction feel like two people working on a problem together?" Joint attention, contingent responses, mutual modeling of each other's understanding — that's the target. Most current systems aren't close, but it's at least the right direction to be pointing.

The maker morning kid didn't consult a tutorial. She got a peer. Forty-five seconds, and the car drove straight.

That's the benchmark.