Who Owns a Child’s Digital Twin?

At a recent parent-teacher meeting, I watched a classroom behavior app assign children “risk” scores that nobody in the room could explain. The stated goal was support. The effect, at least in that moment, was to turn a complicated child into a tidy prediction.

That experience has stayed with me as researchers build something more ambitious: AI models designed to illuminate how a developing brain learns. Used carefully, these systems can function like microscopes. They let scientists test ideas that would be invasive, impractical, or unethical to test directly in children.

But microscopes do not merely reveal. They frame. Someone chooses what enters the field of view, what counts as signal, and what remains outside it.

A model of difficulty—not a duplicate child

A striking recent study used a biologically inspired deep neural network to investigate developmental dyscalculia, a learning disability involving numbers and arithmetic. The model was trained on visually presented arithmetic and designed to resemble parts of the brain’s dorsal visual pathway. Researchers then altered features corresponding to neural patterns associated with dyscalculia and compared the model’s behavior with patterns observed in children (Strock et al., 2025).

Within this artificial system, the results pointed toward difficulty forming number-selective representations in parietal-like layers, rather than a broad failure of attention or working memory. That matters because different mechanisms can produce a similar classroom scene: a child pauses, guesses, or avoids a worksheet. A computational model can help researchers ask which explanation best fits the evidence (Strock et al., 2025).

This is genuinely promising. It is also easy to overstate.

The model is not a copy of a particular child. It does not contain their fear of being called on, the teaching they received, the language used at home, or the strategies they invented to get through yesterday’s homework. “Digital twin” is an evocative phrase, but it can imply a completeness that the science has not earned.

A better description might be a testable model of selected mechanisms. Less dazzling, certainly. More honest, too.

Prediction is not personhood

Another study offers a useful caution from visual neuroscience. Researchers found that neural-network models predicting activity across the visual cortex could be made far smaller while retaining much of their predictive usefulness. Earlier visual areas were easier to capture compactly than higher-level areas, where representations are more complex (Cowley et al., 2026).

The encouraging lesson is that scientific usefulness may not require gigantic, opaque systems. Smaller models can make it easier to inspect which components are doing meaningful work.

The sobering lesson is that matching a neural signal does not make a model equivalent to the brain producing it. A map can preserve the roads needed for a journey while omitting the smell of a bakery, the history of a neighborhood, and who feels safe walking there. Predictive fit tells us that a model captures something. It does not tell us that it captures everything—or even that it has found the only possible explanation.

This distinction becomes ethically urgent when a research instrument leaves the lab. A model built to study mechanisms can become a screening tool. A probabilistic output can enter a school record. A tentative pattern can harden into the sentence, “This is the kind of learner you are.”

The risk is not only that an algorithm may be wrong. It is that adults may stop looking once it produces an answer.

Questions parents can bring into the room

Parents do not need to become machine-learning specialists. Institutions do need to answer ordinary questions in ordinary language:

• What decision will this model influence? Research, screening, classroom placement, and diagnosis are not interchangeable uses.
• Was it tested with children like mine? Ask about language, disability, culture, schooling context, and other forms of developmental variation.
• What information does it collect, and who can access it? Brain, gaze, voice, movement, and performance data can be deeply revealing even when a child’s name is removed.
• Can a family decline without losing support? Consent is thin if saying no carries a penalty.
• Can a qualified human challenge the result? There should be a visible route for correction, context, and appeal.
• When is the data deleted? Childhood records should not quietly become permanent reputations.

If a model’s output is being used to guide diagnosis or treatment, discuss what it means with a pediatrician, psychologist, or appropriately qualified learning specialist rather than treating the score as a conclusion.

Keep the microscope pointed at the system, too

AI may help developmental science separate mechanisms that look identical from the outside. It may allow researchers to test hypotheses without experimenting on a child’s actual brain. Those are meaningful gains.

Yet the ethical question is not simply, “Can this model see a pattern?” It is also, “Who gets to name that pattern, and what happens to the child afterward?”

A humane model should increase adult curiosity, not replace it. It should make support more responsive, not make labels more durable. And it should remain what it is: a deliberately incomplete instrument for understanding part of a developing mind—not a verdict on the person whose mind is still developing.