Every few months, a headline pops up announcing that scientists can now predict intelligence, mental illness, or personality from a brain scan. The implication is usually something like "the future is here, we can see your thoughts in pretty colors." And while the science behind these claims is real, the reality is quite a bit messier than the headlines suggest.
A review in Biological Psychiatry takes a hard look at how well brain-behavior associations actually work in developing children. The short answer: it's complicated. The longer answer involves noisy data, wiggly kids, overstated effect sizes, and the persistent gap between what brain scans can theoretically tell us and what they actually deliver in practice.
The Dream: Brain Scans as Crystal Balls
The promise is genuinely exciting. Large developmental datasets like the ABCD Study now provide thousands of brain scans from children followed over time. Researchers can track how the brain develops across childhood and adolescence while simultaneously measuring cognitive abilities, behavioral traits, and psychiatric symptoms.
In principle, patterns in brain activity should predict real-world outcomes. Maybe certain connectivity patterns in a 9-year-old's brain predict their academic performance at 12. Maybe activity patterns during a task reveal something about their risk for depression. Maybe we can catch problems early and intervene before they become serious.
That's the dream. Both task-based fMRI (where kids perform specific activities in the scanner) and resting-state fMRI (where they just lie there doing nothing) offer windows into brain function. Put it all together, and you should be able to build models that connect brain organization to behavior.
Should. That's the key word doing a lot of heavy lifting.
The Nightmare: Children in MRI Scanners
Here's a fundamental problem with scanning kids' brains: kids are kids. They move. They get bored. They randomly press buttons because buttons are fun to press. They fall asleep. They have existential crises about whether they remembered to bring their snack.
MRI is exquisitely sensitive to motion. Move your head a few millimeters, and you've corrupted the data. Adults can usually hold still for an hour. Children... not so much. Movement artifacts can introduce systematic biases that even sophisticated correction methods struggle to fully address.
Beyond motion, there's the problem of attention. Task-based fMRI requires participants to actually do the task. If a 10-year-old spaces out for half the scan, the data from that period doesn't reflect what you're trying to measure. And good luck knowing exactly when the spacing out happened.
These aren't minor technical annoyances. They're fundamental challenges that limit how much signal you can extract from developmental neuroimaging data.
Effect Sizes: The Disappointing Truth
Here's the part that really stings. Many of the brain-behavior associations that made headlines have turned out to be overstated. When researchers use proper methodology with large samples and appropriate statistical controls, the effect sizes shrink. Sometimes dramatically.
That exciting finding that brain connectivity patterns predict IQ? When you do it right, maybe it explains 2% of the variance in intelligence. That's not nothing, but it's also not exactly crystal ball territory. You'd be better off predicting IQ from a five-minute conversation than from an expensive brain scan.
This doesn't mean brain-behavior associations don't exist. They do. It just means they're modest, not magical. The brain scan is not reading your child's future. It's picking up on very subtle statistical patterns that explain a small fraction of individual differences.
For clinical applications, small effect sizes create real problems. If your prediction model only explains 2% of who will develop depression, it's not going to be particularly useful for making decisions about individual patients.
The Generalization Problem
There's another complication. Results from one dataset often don't replicate in another. A model trained on suburban California kids might not generalize to kids from different backgrounds, regions, or socioeconomic contexts.
This matters a lot for any practical application. If your brain-behavior model only works on populations similar to the one it was trained on, it's not really a general finding about human brain development. It's a finding about that particular sample.
The field is slowly getting better at testing generalization explicitly and building more diverse datasets. But we're not there yet, and overgeneralizing from limited samples remains a persistent problem.
Doing It Better
The review isn't all doom and gloom. It highlights emerging approaches that could improve the situation. Better acquisition protocols that minimize motion artifacts. More sophisticated analysis methods that can handle the complexity of developmental data. Machine learning approaches that extract multivariate patterns rather than relying on simple correlations.
There's also a growing recognition that the behavioral measures matter as much as the brain measures. If you're trying to predict a behavior that's poorly measured in the first place, no amount of sophisticated neuroimaging will help. Garbage in, garbage out, but with a really expensive scanner.
Kids Don't Develop in a Vacuum
Perhaps the most important point in the review is that brains don't develop in isolation. Children grow up in families, neighborhoods, schools, and cultural contexts. Socioeconomic status, stress exposure, educational opportunities, and countless other environmental factors influence both brain development and behavioral outcomes.
Treating the brain as if it were the sole determinant of behavior ignores this obvious reality. Future work needs to integrate environmental and contextual factors rather than pretending that a brain scan can tell you everything about a child's trajectory.
The brain-behavior associations in developing children are real. The brain really does matter for cognition and behavior. But the relationships are more complicated than a headline can convey, and extracting reliable, generalizable predictions from brain scans remains a work in progress.
Reference: Rosenberg MD, et al. (2025). Task and Resting-State Functional Magnetic Resonance Imaging Modeling of Brain-Behavior Relationships in Developmental Cohorts. Biological Psychiatry. doi: 10.1016/j.biopsych.2025.09.012 | PMID: 41043534
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.