Your Kid Is the Beta Test

There's a particular kind of cognitive dissonance that comes from holding two things in your head at once: this is technically correct, and this is quietly wrong.

I spent part of last week reading through a draft AI literacy curriculum being piloted in a neighboring school district. The materials were well-designed. Accurate. Age-appropriate in the narrow sense. They explained machine learning clearly, covered bias (briefly), and taught kids how to prompt and evaluate AI outputs. Every lesson was organized around the same implicit premise: AI is a tool that you use.

Not a system that was designed by someone with choices and values. Not an intervention that is shaping you as you interact with it. A tool.

I kept flipping pages looking for the lesson that asked: Who built this, and what did they optimize for? It wasn't there.

I came away uncomfortable in a specific way — not because the curriculum was wrong, but because it was teaching children to learn about AI without ever learning to question it. And it was doing this while those children's brains are in the most sensitive, most irreversible windows of their development.

The Timing Is the Thing

Here's what the neuroscience tells us that most technology policy conversations quietly ignore: children's brains are not small adult brains. They are systems in the middle of a timed sequence of developmental events, and the sequence matters enormously.

A 2025 paper in Neuropsychopharmacology synthesizes evidence that critical periods in the human brain unfold hierarchically — from primary sensorimotor regions first, to higher-order association cortices that govern reasoning, planning, and social cognition later ("Investigating Hierarchical Critical Periods," 2025). Different psychological processes are most sensitive to environmental input at different developmental times. Disrupting that sequencing — exposing the wrong system to the wrong input at the wrong moment — is implicated in youth-onset psychiatric conditions including schizophrenia, depression, and autism.

This isn't metaphor. The timing of what enters a developing brain matters in ways it simply doesn't for a fully-formed adult brain. When we put a five-year-old in daily contact with an AI system that responds to them with consistent warmth, never loses patience, and never admits uncertainty, we are not doing the same thing to them that we're doing to their teacher who also uses the tool. We're doing something during a window we can't reopen.

Early Environments Don't Just Shape Development — They Alter Its Machinery

The epigenetics research makes this point even starker. A comprehensive 2025 review in Neuropsychopharmacology examines how early-life environments become stably encoded in the epigenome, permanently altering the expression of plasticity-regulating genes and reshaping the timing and magnitude of sensitive period windows themselves ("Epigenetic Regulation of Brain Development," 2025). The brain's experience-dependent architecture is not just sculpted by what inputs arrive. It is shaped by how the molecular machinery of plasticity is calibrated by early experience.

Put plainly: the environment a child grows up in doesn't just affect what they learn. It affects how their brain will be capable of learning for the rest of their life.

We are asking whether AI tutors raise test scores. We are not asking what consistent interaction with AI systems does to the developing neural machinery that governs curiosity, epistemic trust, tolerance for ambiguity, and the willingness to interrogate authority. Those are not small questions. They are, arguably, the whole ballgame.

What the Ethics Review Board Actually Found

In March I spent a week on a regional ethics board evaluating AI tools proposed for use in public school curricula. I left disturbed by something specific: almost none of the submitted systems had any documentation of how they'd been tested on children. They had user testing data, in many cases. Pilot results, sometimes. Validation studies in adult populations, occasionally. But children — the population whose brains were actually going to be shaped by these interactions — almost nothing.

This is not a hypothetical gap. This is the actual state of deployment. We are putting AI systems into classrooms in ways that would never pass clinical review in any other context involving children's development.

In medicine, we learned this lesson at enormous cost. The history of pediatric pharmacology is littered with adult-dosing assumptions catastrophically misapplied to children — not because the drugs worked differently in principle, but because dosing, metabolism, and developmental timing interact in ways that can't be assumed across age groups. The assumption that "safe for adults, probably fine for kids" is not a policy. It is a wish.

Even the Clinical AI Advocates Are Worried

The same caution applies in mental health contexts, where AI is being deployed with adult patients and where at least some regulatory pressure to demonstrate clinical validity exists. A 2025 review in Science surveying AI's potential to transform mental health care is genuinely enthusiastic about what's possible — digital phenotyping, personalized treatment selection, population-level surveillance — but is equally blunt about the unresolved challenges: algorithmic bias, privacy risks, lack of explainability, and the near-total absence of clinical validation for most deployed systems ("Transforming Mental Health Research," 2025).

A corresponding 2025 perspective in Nature Computational Science makes the case for grounding AI tools in mechanistic brain models rather than purely data-driven black-box approaches, precisely because black-box systems fail to generalize across individual differences ("Transforming Psychiatry," 2025). Children are, biologically speaking, a population defined by the particularity of their individual developmental timing. A system trained on adult data and validated against adult norms is, with respect to children, essentially ungoverned.

These are the arguments being made in the adult clinical mental health literature, where a regulatory framework — imperfect, slow, often captured — at least nominally exists. In educational technology, that framework barely has a pulse.

The Consent Question Nobody Is Asking

Consent frameworks exist not just to protect people from immediate harm, but to preserve their capacity to be authors of decisions that affect them.

Children can't consent to being part of the longitudinal experiment that is AI deployment in schools. Their parents can consent on their behalf — and sometimes do, in the form of district-adopted technology agreements that run to forty pages of legal boilerplate. But there's a structural problem: we don't know what we're asking them to consent to. We haven't done the child-specific testing. We don't have long-term outcome data. The consent form is, in effect: Please sign to acknowledge that your child will participate in a study we haven't fully designed yet.

The AI literacy curriculum I reviewed last week framed this beautifully — unintentionally. One lesson asked students to evaluate whether an AI-generated answer was correct. A reasonable exercise. But there was no lesson asking students to evaluate whether the AI's values — what it was optimized for, what tradeoffs its designers made — were ones they agreed with. The curriculum taught children to be good users of AI, not critical thinkers about it.

And it is doing this during the window when the habits of epistemic mind are being formed. Not the facts they'll carry. The posture they'll take toward knowledge for the rest of their lives.

What Responsible Deployment Would Actually Look Like

I try not to be the person who just says "but have you considered the ethics?" and walks away. So:

Child-specific testing should be required before educational AI is deployed at scale — the same way pediatric clinical trials are required before approving a drug for children. The testing should include not just immediate outcomes (did the child perform better on the assessment?) but developmental outcomes tracked over time: what happens to their tolerance for ambiguity, their trust in human teachers, their willingness to be publicly wrong?

AI literacy curricula should include lessons that teach children to be critical of AI systems, not just competent with them. "Who built this?" and "What were they trying to maximize?" are good questions. So is "What does this system not know how to tell me?" A child who can prompt an AI but cannot interrogate one has been handed a tool without a manual.

And wherever children are old enough, they should be included — not as passive subjects of AI deployment, but as participants in conversations about it. Their intuitions about what feels manipulative, what feels genuinely helpful, and what feels quietly wrong are not noise. They are data we are currently discarding.

---

The most important question in AI has never been how smart we can make it. It's who it affects, and whether anyone asked them. We keep discovering the hard way that "who it affects" includes people who are still in the middle of becoming who they'll be.

The window for getting this right is, like all developmental windows, finite.