Social Learning Built Human Cognition. AI Is Taking Notes.

Here is a small experiment that tells you a lot about how humans learn. In the early 2000s, developmental psychologist Patricia Kuhl ran a study with American infants learning Mandarin phonemes. One group received live sessions with a Mandarin-speaking adult — full eye contact, contingent responses, warmth. Another group received the same content on a video screen. A third heard only audio. After a dozen sessions, the live-interaction group had learned to distinguish the new sounds. The video and audio groups showed nothing. The content was identical. The relationship was not.

This finding has been replicated, extended, and refined in the two decades since. What we now understand is that social interaction doesn't just motivate learning — it neurobiologically primes for it. And that understanding raises a question I've been sitting with more than usual lately, after a week serving on a regional ethics review board evaluating AI tools proposed for public school curricula: when AI systems begin to mimic the social signals that make human learning possible, what are we responsible for?

The Neural Architecture of Learning Together

Joint attention — the act of two beings orienting to the same object or event — sounds unremarkable. It is, in fact, one of the most cognitively sophisticated things a human infant does. By nine months, most infants can follow a caregiver's gaze, point to share interest, and coordinate attention with another person around something in the world. This capacity is widely considered foundational to language, theory of mind, and cultural transmission. It is, in a meaningful sense, one of the things that makes human civilization possible.

What's new is how precisely we can trace its neurological roots. A 2024 study mapped the whole-brain connectivity patterns in infants aged 8 to 15 months, looking for neural signatures that predicted joint attention behavior (Grossmann et al., 2024). The dominant connections lay within the default network and its interaction with the ventral attention network — regions long associated with social cognition and self-referential processing. More striking: this same joint attention connectome, established before the first birthday, predicted Theory of Mind ability at ages 2 to 5. The architecture for understanding other minds appears to be laid down before language even arrives.

But this wiring doesn't activate in isolation. A landmark study from the University of Washington's Institute for Learning and Brain Sciences used magnetoencephalography to record brain activity in 5-month-old infants during live face-to-face social interaction versus a non-social control (Bosseler et al., 2024). Neural theta activity in right-hemisphere attention and sensorimotor regions during that live social exchange predicted language development at five follow-up time points stretching from 18 to 30 months — more than two years later. Not mere language exposure, mind you. Social interaction specifically. The "social ensemble" of eye contact, infant-directed speech, and contingent responses primes the brain to learn. Remove the contingency, and the priming signal weakens.

This is why video screens didn't work in Kuhl's study. It wasn't resolution or audio quality. It was whether the other agent responded to you.

Contingency Is the Signal

Developmental robotics researchers have been testing this principle from the opposite direction: if contingent responsiveness is what makes an agent socially formative, can robots produce the same priming effect? A 2024 scoping review synthesized findings from dozens of studies using social robots as experimental partners with children aged 2 to 35 months (De Greeff et al., 2024). The conclusion was nuanced but telling. Infants and toddlers can pay attention to, learn from, and read social signals from robots — but only when those robots produce behaviors that are interactive and contingent. Static or pre-scripted robots didn't cross the threshold. The children who treated robots as genuine social agents were those who received robots that genuinely responded to them.

This is simultaneously a scientific finding about social development and, I think, a design specification for AI systems. If contingent responsiveness is the key variable — if it is what activates the neural priming for learning — then engineers building AI tutors for children have been handed a developmental psychology roadmap.

The question is whether they are following it responsibly. And whether anyone is checking.

When Infant Learning Becomes a Design Template

The research linking infant social learning to AI architecture has moved quickly. A 2025 study from the Weizmann Institute trained AI models on social prediction tasks in the manner that infants approach them: first learning to detect animacy and attribute goals, then building toward more complex social inferences (Treger and Ullman, 2025). Models that followed the infant scaffolding — concept-first, incremental, socially grounded — dramatically outperformed standard deep learning approaches on training efficiency, accuracy, and generalization to novel actors and scenarios. Standard deep learning, it turns out, skips the scaffolding that human development relies on. Mimicking the developmental sequence made the machines better at the task.

Meanwhile, computational developmental psychology is formalizing what was previously intuitive. A 2025 Annual Review paper from Yale's Computational Cognitive Development Lab models children's Theory of Mind acquisition using Bayesian frameworks, showing that parent-child conversation and social experience systematically shape children's ability to reason about other minds (Jara-Ettinger, 2025). Social interaction, in this account, is not background noise in development — it is the primary signal. The relationship between child and caregiver is, in some sense, the training data.

If that framing sounds uncomfortably close to how machine learning researchers talk about dataset curation, you are not wrong to notice.

The Question Nobody Asked

I recently spent a week reviewing AI tools proposed for public school curricula through a regional ethics review panel. I came away with a specific kind of unease that I want to try to articulate carefully, because I think it matters beyond the particular submissions I reviewed.

Several of the systems under evaluation were, in one way or another, designed to engage children's social learning instincts. They mimicked contingent responses. Some produced avatar-based affect or simulated turn-taking in ways that resembled — functionally, at least — the social ensemble that Bosseler and colleagues show is neurobiologically formative. And almost none of them had documentation showing how they had been tested on children specifically. The testing populations, where disclosed at all, were adults.

This is not merely a methodological gap. A 2025 article introducing the interdisciplinary journal AI, Brain and Child maps both the transformative potential and the critical risks of AI during sensitive developmental windows, arguing explicitly for longitudinal research into how AI systems shape developing minds during the first 25 years of brain development (Springer Nature / ABC Journal Editorial Team, 2025). What the authors describe as urgently needed is essentially what those product submissions were missing: evidence that the systems had been evaluated in the populations they were designed to reach.

The history of technology deployment in children's spaces offers ample precedent for why this matters. Television was introduced into American classrooms without systematic studies of its developmental effects. The same is broadly true of smartphones. In each case, the technology moved faster than the developmental science, and children served as the de facto test population. The difference now is that we have the developmental science. We know that infants' brains respond differently to contingent versus non-contingent social agents. We know that joint attention is a neurological precursor to theory of mind. We know, from Treger and Ullman (2025), that infant-style scaffolding produces meaningfully different outcomes than standard architectures.

What we don't yet have is a policy framework requiring AI systems designed for children to demonstrate that they've been evaluated using standards that developmental science would demand.

Genuine Complexity

Here I want to resist the temptation of easy answers, because the genuine complexity deserves respect.

On one hand, the evidence is fairly clear that social interaction primes human learning in ways that asynchronous instruction does not. If AI tutors can provide more children with access to contingent, responsive educational support — particularly children in under-resourced schools who may not have consistent access to skilled human teachers — that is a serious potential benefit. Dismissing it on the grounds that the technology is imperfect or unproven would be its own kind of irresponsibility.

On the other hand, children are a specific and vulnerable population. Their brains are, in a non-metaphorical sense, still being built. The mechanisms that social AI is designed to engage — contingency detection, joint attention, social imitation — are the same mechanisms that developmental researchers study because they are foundational to cognitive and social development. Engaging those mechanisms through artificial agents, at scale, during sensitive developmental periods, without documentation of effects, is not a neutral act.

Several things seem warranted regardless of where one lands on the broader debate. First: AI systems designed for children should be required to document their testing methodology with child populations — at minimum before adoption in public school settings. Second: developers should be expected to engage with developmental science not as a marketing claim but as a methodological standard. Third: the longitudinal question — what happens to children who spend significant time with socially responsive AI during early development — needs to be studied proactively, not retroactively.

School administrators, curriculum specialists, and policymakers considering AI tutoring tools should consult with developmental psychologists as part of the adoption process, not after. And if you're navigating AI regulation or data privacy compliance in educational settings, a technology policy attorney can help clarify the legal requirements — this area is evolving fast.

The Design Choices That Echo

Here is what I keep coming back to. The developmental science tells us that social learning is not incidental to how children become cognitive agents — it is foundational. Joint attention, imitation, contingent responsiveness: these are the mechanisms through which culture transmits itself across generations, through which children build their models of the world and of other minds. They evolved over hundreds of thousands of years. They are among the things that make us distinctively human.

We are now deliberately engineering AI systems to interface with those mechanisms. That is a significant design choice. It is also, in some respects, a genuinely exciting one — the findings from Treger and Ullman (2025) suggest that understanding infant learning could improve AI performance, and the neurobiological work from Bosseler et al. (2024) and Grossmann et al. (2024) offers a real window into how to design systems that support rather than distort natural development.

But "should we?" has to accompany "can we?" at every step. And right now, across the AI-in-education space, there is considerably more evidence of the latter being asked than the former.

The design choices we make now — about which social signals AI systems mimic, how those systems are tested, who reviews their developmental appropriateness, what counts as evidence of benefit and harm — will shape how a generation of children develops. That's not hyperbole. That's what the developmental science tells us. The social ensemble that a 5-month-old brain responds to isn't arbitrary. It has been selected for over evolutionary time. When we build artificial systems to simulate it, we should at least understand what we're touching.