The Best Brain Model Might Be Tiny

A rainy morning recently found me reorganizing my bookshelves. By author, I could find a scholar quickly but lose an entire subject. By subject, interdisciplinary books became bureaucratic emergencies. Every arrangement revealed something—and hid something else.

AI models used in developmental research work the same way. They can organize a messy stream of looks, choices, movements, or brain activity so that a pattern becomes visible. But the model’s success depends on what researchers asked it to preserve. This is why calling AI a “microscope” is useful only if we remember that microscopes have lenses, settings, and someone deciding where to point them.

The most revealing model may not be the biggest one. It may be the smallest model that still captures the pattern scientists care about.

Prediction is a clue, not an explanation

Imagine a child repeatedly choosing between two games. Their choices may look inconsistent: they return to a game that just disappointed them, suddenly switch after a streak of success, or explore when an adult expects them to exploit what they know. A researcher can hand-build a theory for each tendency. Alternatively, a compact learning model can search for a strategy that makes those choices less mysterious.

Li and colleagues trained tiny recurrent neural networks on reward-learning behavior from people and other animals. The compact networks predicted behavior more accurately than several traditional hand-crafted models and exposed interpretable strategies behind quirks and individual differences (Li et al., 2025). Their achievement was not making a miniature person. It was building a manageable hypothesis generator.

That distinction matters. If a model predicts when someone will switch choices, it has found regularity. It has not necessarily discovered why the person switched. The same outward behavior could reflect curiosity, fatigue, uncertainty, or a misunderstood instruction. Prediction narrows the puzzle; it does not get to declare the puzzle solved. Science remains annoyingly employed.

Compression can reveal the useful machinery

Large neural networks often contain many routes to a successful answer. That helps performance, but it complicates explanation. When nearly everything participates, “the model learned it” is not much better than saying “the brain did it.” Technically true. Spectacularly unhelpful.

Cowley and colleagues compressed neural-network models used to predict activity along the visual cortex while preserving much of their ability to match neural responses. Models of early visual areas were easier to compress than models of later areas, where representations are more complex (Cowley et al., 2026). Compression therefore did more than save computing power. It helped separate machinery that mattered for brain prediction from machinery that was largely redundant.

For developmental neuroscience, this is a powerful ideal: use machine learning not merely to fit children’s data, but to identify the leanest set of computations that could produce the observed developmental change. A smaller model is not automatically correct. It is simply easier to question—and good science depends on models that can survive awkward questions.

Children are not just smaller benchmarks

The comparison becomes especially useful when children and machines face the same carefully controlled task.

Ayzenberg and colleagues showed preschoolers ages 4–6 and leading vision models objects with local visual information disrupted. The children recognized objects rapidly and remained robust under alterations that can unsettle machine vision, although child and model performance also overlapped in meaningful ways (Ayzenberg et al., 2025).

The interesting result is not “children win.” Benchmarks are not playground scoreboards. The value lies in the shape of each learner’s errors. If a child recognizes a strangely altered cup while a model fails, researchers can ask what information the child preserved: overall form, likely function, prior physical experience, or some combination. An AI failure becomes a contrast dye, making a feature of human learning easier to see.

But the reverse is also important. A model can match a child’s answer for the wrong reason. Agreement in output does not guarantee agreement in representation. Two students can circle the same answer; only one may understand the question.

What parents should take from an “AI reads the brain” headline

First, ask what the system actually predicted. A gaze direction, a category label, and later learning are very different targets. “Decoded learning” can conceal several definitions wearing one impressive trench coat.

Next, look for the task and the comparison group. A pattern found in a controlled experiment may illuminate how learning works across children without offering a meaningful interpretation of one child at home.

Also ask whether researchers tested unfamiliar examples. Children routinely encounter new accents, odd drawings, cluttered rooms, and instructions delivered while someone is looking for a missing shoe. A model that works only on data resembling its training set may be detecting the laboratory’s organization rather than the child’s understanding.

Finally, treat individual predictions as probabilities, not identities. A model can help scientists notice variation. It should not turn a developing child into a fixed label, especially when behavior changes with context, relationships, sleep, language, and experience.

AI’s best role in developmental science is not to pronounce what is inside a child’s mind. It is to make hidden patterns discussable, testable, and easier to challenge. Sometimes the clearest lens is a compact model that shows its workings. Sometimes it is a benchmark that fails exactly where a preschooler does not.

The microscope earns our trust not by sounding certain, but by helping us see what question to ask next.