AI Can Now Predict ADHD in Four-Year-Olds: Is it a Gift or a Trap?
A new Duke study trained an AI tool on 140,000 children’s medical records to flag ADHD risk years before diagnosis.
When researchers at Duke University announced last week that they’d built an artificial intelligence tool that can flag ADHD risk in children years before any diagnosis is made, the headlines wrote themselves. Faster diagnosis. Earlier support. Fewer kids slipping through the cracks.
Published in Nature Mental Health on April 27th, the Duke team trained their model on routine electronic medical records from more than 140,000 children, teaching it to recognize the combinations of developmental, behavioral, and clinical events that tend to show up years before pediatricians eventually write the letters A-D-H-D in a chart.
The model works. It’s accurate across sex, race, ethnicity, and insurance status, and it performs well in kids age 5 and older. Senior author Matthew Engelhard, M.D., Ph.D., was careful to emphasize what the tool isn’t:
“This is not an AI doctor,” he told reporters. “It’s a tool to help clinicians focus their time and resources, so kids who need help don’t fall through the cracks or wait years for answers.”
I read all of that, and I thought, well, that depends entirely on what we think “the cracks” are.
Here’s the thing about a tool like this. It’s morally neutral. It just finds patterns. Whether it ends up helping a generation of Hunter kids or harming them depends almost entirely on what the adults in the room believe about ADHD in the first place.
And the adults in the room, meaning the medical system, the schools, the insurance companies, the pharmaceutical industry, and the average pediatrician, have been working from a “deficit” model for almost ninety years.
Consider the importance — and impact on kids’ lives — the fork in the road this study could present.
If you accept the standard model, the one that says ADHD is a defective attention system that needs to be corrected, then an AI tool that flags it earlier sounds wonderful. Catch the broken brain at four instead of nine. Start the stimulants sooner. Get the kid into special education before he falls behind. Reduce the risk of the long list of bad outcomes supposedly waiting for an unmedicated ADHD kid. Save five wasted years.
But if you’ve been reading my work for any length of time, you know I see this differently. I’ve been arguing for thirty-three years that ADHD isn’t a defective attention system. It’s an evolutionary inheritance.
The traits that look broken in a fluorescent-lit second-grade classroom are the same traits that kept hunting bands alive for half a million years. Hunters scan widely, react quickly, hyperfocus on what’s interesting, get bored fast with what isn’t, take risks, notice patterns nobody else sees, and recover from failure faster than any Farmer would believe possible.
If that’s the model in your head, then the Duke AI suddenly looks completely different.
Imagine an AI tool that flags four-year-olds with Hunter brains. Imagine getting that information at age four. Now imagine two completely different responses to the same flag.
Response one is the current system on autopilot. Earlier evaluation. Earlier diagnosis. Earlier referral to a psychiatrist. Earlier prescription. Five extra years of being told there’s something wrong with you. Five extra years of medication. Five extra years of being managed instead of understood. The flag becomes a label. The label becomes an identity. The kid grows up believing he’s the broken one.
Response two starts with the same flag and ends somewhere else entirely. The pediatrician says:
“Your child appears to have a Hunter brain. Here’s what that means. Here’s what tends to work for these kids, and what tends to backfire. Let’s talk about how to set up his life so the wiring works for him instead of against him.”
The parents go home and start designing the right environment. They understand why he won’t sit still through circle time, and stop fighting it. They make sure he gets enough physical movement, enough sleep, enough novelty, enough engagement with things that genuinely interest him.
They find a school that doesn’t try to flatten Hunter kids into Farmer shapes. They build a life around the Hunter instead of trying to bend the Hunter into a life that was never built for him. Medication may still be part of the picture, but it’s one tool in a toolbox, not the entire response.
Same kid. Same flag. Same AI prediction. Two completely different futures.
There’s a deeper issue in the technology itself that I want parents and clinicians to take seriously. The Duke model didn’t learn to identify Hunter brains; that’s not what it was trained to do. It was trained on the diagnostic decisions of the existing system, learning the pattern of medical events that historically precede an ADHD diagnosis in our current healthcare environment.
Which means the AI isn’t really predicting ADHD. It’s predicting whether your particular child is on a trajectory to be processed through the existing diagnostic machine. The model sees future patients, not future successful Hunters. Whether being a future patient is a good outcome or a bad one depends entirely on what that machine does once it has them.
There’s something else worth noticing, and it has a quiet beauty to it.
An AI that scans massive datasets looking for subtle patterns nobody else has seen, that picks up signal where everyone else sees noise, that recognizes a constellation of small events that together tell a larger story, is doing exactly what a Hunter brain was built to do!
We’ve built a digital tracker. The question is whether the people deploying it are tracking with the eye of a Farmer who wants to fence Hunters in, or with the eye of someone who wants to find them and set them up for the lives they were actually built for.
A personal note. I was a behavior problem early on; I still remember Mrs. Clark in 2nd grade saying, “Tommy, even a fish wouldn’t get caught if it kept its mouth shut,” and, “An empty wagon always rattles.” School was hard for me until I was admitted into a gifted kids class when I was eight (Eisenhower launched that program in response to Sputnik), and most of the adults around me had concluded that the problem was me.
If a tool like the Duke model had existed when I was four and had flagged me, I’m honestly not sure whether that would have made my childhood better or worse. The flag itself is just information: what would have made the difference is what came next.
Did the adults around me have a framework for understanding what they were looking at? Did anyone tell my parents that the kid bouncing off the walls might grow up to build businesses, save lives, write books, and host America’s number one progressive national radio show? Did anyone help them build a life around how I’m wired?
Or would early identification just have meant earlier medication, earlier stigma, and earlier conviction on my part that I was the broken one?
That’s the question every parent should ask if their pediatrician starts using a tool like this. Not “is the AI accurate,” because it probably is. The question is, “What do you, doctor, plan to do with the information?”
If the answer is “earlier evaluation and likely earlier medication,” that’s one path. If the answer is “earlier conversation with you about how to set this child up for the life he’s wired for,” that’s a very different one.
We’re about to have a lot more information about Hunter kids a lot earlier than we ever have before. That information can liberate them, or it can lock them in. The technology has arrived. The question is whether our understanding has caught up to it.
If you want to keep these conversations going, subscribe and share this with the parent of a young Hunter who needs to hear it. The tools are getting more powerful by the year. We need to make sure the framework we use them inside of is worthy of the kids we’re using them on.