Children Learn to Sort People. AI Never Stopped.

In 1947, psychologists Kenneth and Mamie Clark sat across from Black children as young as three years old and asked them to choose between a white doll and a Black doll. Most children preferred the white doll — calling it "nice," describing the Black doll as "bad." The Clarks understood what they were documenting: children were absorbing society's racial hierarchy before they could read or write. They were, in the language of our current moment, being trained on biased data.

The difference, of course, is that those children were also growing up. They were developing the capacity to be taught otherwise.

That developmental arc — absorb the culture, then interrogate it — is something we have almost entirely failed to build into the AI systems that are increasingly mediating those children's educations, social lives, and sense of self.

How Children Learn to Sort

Social categorization isn't a failure of development. It's a feature of it. Infants detect race, gender, and age distinctions within the first months of life — not because they've been taught prejudice, but because the brain is a pattern-recognizing organ, and social patterns are among the most consistent in a child's environment. By age three, children show measurable in-group preferences. By five, they've begun attaching traits to those categories: who is "smart," who is "nice," who is "in charge."

This is not destiny. One of the most hopeful findings in developmental psychology is that children are also sophisticated moral reasoners — and relatively early ones. By school age, they understand fairness principles with genuine conviction. Given the right environment — the right teachers, stories, peer relationships, explicit conversations about why stereotypes are wrong — children can revise their social category associations. Not perfectly, not quickly, and not always. But the mechanism for update is built into the developmental process itself.

The question is whether we've built anything like that mechanism into the systems we're now deploying at scale.

What AI Inherits

A useful reframe comes from Yiu, Kosoy, and Gopnik (2024), who argue that large language models are fundamentally cultural transmission engines. They are extraordinarily good at faithfully replicating the patterns encoded in their training data — not transforming, critiquing, or selectively discarding those patterns, but transmitting them with high fidelity. As the authors note, models "can produce overimitative text that mirrors stylistic patterns without understanding causal structure."

This is worth sitting with. Human cultural transmission has always encoded social hierarchies — in who tells stories, whose stories get written down, who appears in textbooks as the default protagonist and who appears as a footnote. When those hierarchies are replicated billions of times across the internet, and an AI is trained on that text, the model isn't just learning language. It's learning a social world in which those hierarchies are the baseline assumption.

Children absorb this same environment. But children are selective imitators. They copy intentional actions, discard failures, and can generate novel solutions when the template doesn't fit. According to Yiu, Kosoy, and Gopnik (2024), children "specifically copy intentional actions, discard failed attempts and causally irrelevant steps" — a selectivity that fundamentally shapes which cultural patterns get reinforced and which get revised. AI systems, optimized for distributional fidelity, do none of this. They replicate more comprehensively and less critically than any individual human ever could.

The Surface Without the Depth

There's a subtler problem underneath the statistical one. Mahowald et al. (2024) draw a principled distinction between formal linguistic competence — mastery of the statistical structure of language — and functional linguistic competence — using language to reason, plan, and communicate about the world. LLMs, they argue, excel at the first and fail systematically at the second. Critically, the brain's language network handles formal competence, while reasoning and world knowledge are computed by separate, non-linguistic systems — systems that don't have a direct analog in current AI architectures.

Applied to social categorization: a language model can reproduce, with uncanny fluency, the surface patterns of how humans talk about race, gender, class, and disability. It can complete sentences, fill in associations, and generate descriptions that reflect (or amplify) the categories and hierarchies in its training data. What it cannot do is understand what any of this means at the conceptual level — why these categories exist, what purposes they serve, who benefits from them, and why they cause harm.

This asymmetry creates a particular kind of danger. A person who has absorbed stereotypes can also, in principle, be helped to understand why those stereotypes are wrong — because the capacity to learn social categories and the capacity to reason about their moral significance operate within the same mind. An AI system that replicates social categories at the level of statistical pattern has no corresponding conceptual apparatus in which a correction can take hold. You can adjust the outputs. But what is actually changing underneath remains genuinely open.

Can Thinking Help?

There's active work on whether prompting AI systems to reason explicitly about their outputs can improve their handling of bias. This connects to broader questions about what Lombrozo (2024) calls "learning by thinking" — the capacity to gain new understanding through internal processes like explanation, analogy, and deliberate step-by-step inference. When LLMs are asked to reason through a problem before answering, they sometimes reach more accurate conclusions, even correcting errors they would otherwise make.

Lombrozo is careful not to overclaim. She notes that LLMs "make the same characteristic errors as humans during LbT" — confidently wrong conclusions from plausible-sounding reasoning chains. And she cautions against conflating the outward form of reflective reasoning with the underlying mechanism that makes such reasoning genuinely productive in humans.

The same caution applies to bias. Chain-of-thought prompting can sometimes surface less biased outputs. Whether this represents genuine recalibration of the model's underlying representations, or a sophisticated surface performance of having recalibrated them, is not something we currently have good tools to determine. This question matters enormously when you're deploying these systems in contexts that shape real decisions — or where children are learning how the world works.

A Historical Parallel That Should Make Us Uncomfortable

We've been here before, though at smaller scale. The battles over school textbooks in the mid-twentieth century — which histories to teach, whose heroes to elevate, what languages to recognize in which classrooms — were, at their core, fights over which cultural patterns would be transmitted to children and in what form. Those were slow, contested, imperfect processes. They required decades of advocacy, litigation, and deliberate design before the transmission channels began to shift.

We are making analogous decisions now, often much faster and with far less democratic deliberation, about AI training datasets: which voices, which texts, which social patterns get included in the data that shapes how these systems understand and categorize the world. The institutions making those decisions are largely private. The communities bearing the costs when they go wrong are largely not.

The parallel isn't perfect — algorithms aren't textbooks, and training datasets aren't curricula. But the underlying question — whose cultural patterns get treated as the default, and what recourse exists when that defaults causes harm — is recognizably the same. And our institutional capacity for course-correction appears, if anything, weaker than it was in 1965.

What Actually Needs to Happen

Algorithmic fairness work — debiasing training data, auditing outputs, adversarial testing — is necessary. It is not sufficient, and it shouldn't be mistaken for the end of the problem. The harder question is structural: which organizations have the power to decide what enters a training dataset, and which communities bear the costs when those decisions go wrong?

Children don't develop better social cognition simply by being exposed to less biased information. They develop it through relationships, explicit moral instruction, conflict, repair, and the slow work of building the capacity to see other people as fully human. AI systems aren't going through any of that. They're having their outputs adjusted and being asked to approximate fairness.

For educators, researchers, and practitioners deploying these systems in high-stakes contexts — hiring, lending, school placement, criminal justice — the honest framing is this: the systems you are using have inherited human patterns of social categorization, do not understand them in any deep sense, and can at best produce outputs that resemble fair reasoning without the mechanism behind it. That's not a categorical argument against deployment. But it is an argument for maintaining meaningful human oversight, investing in ongoing independent auditing, and resisting the comfortable belief that getting the aggregate statistics right means the harder problem is solved.

The Clarks knew something we keep relearning: you can measure bias in a three-year-old. The harder work is building the conditions under which that child — or that system — can grow beyond it. We spent decades, imperfectly, trying to build those conditions for children. We have barely begun to ask what the equivalent would look like for AI.