A force for good: What a teenager's eye scan taught me about AI's better future

Most weeks, the AI stories that land in my feed are about what these systems might take from us. Jobs, privacy, trust, the energy it takes to run a single query. I read them because I have to, because understanding the risks is part of building anything responsibly. But every so often a story comes along that is good news without any asterisk, and I want to give it the same attention we give the bad ones.

This is one of those stories.

Edward Kang is seventeen. He's a senior at Bergen County Academies in New Jersey, and three years ago he was reading research papers for a school project when he came across a study out of the Chinese University of Hong Kong. The researchers had found a link between retinal images and autism. Kang found the idea both fascinating and slightly strange, that something as ordinary as an eye scan could tell you anything about what's happening inside a brain.

It's a fair reaction, until you remember that the retina isn't really a separate organ at all. It's an outgrowth of the central nervous system, the one part of the brain you can photograph without going anywhere near a scalpel. If a difference in brain development shows up anywhere in the body that a camera can reach, the retina is the obvious place to look.

So Kang taught himself the machine learning to look properly. The result is called RetinaMind, and it's built on the kind of retinal images an optometrist takes during a completely routine eye exam. The neural network he trained picks up patterns in those images too fine for a clinician to see with the naked eye, and uses them to flag whether a scan looks neurotypical or shows signs associated with autism or ADHD. In testing, it's been accurate around 89 percent of the time. Rather than handing back a bare verdict, the tool also produces a heat map of the retina, highlighting the regions that shaped its result, so the reasoning stays visible rather than locked inside a black box.

What impresses me most is that Kang didn't stop at building a classifier. He went on to build a retinal cell model of autism and used it to flag around a dozen candidate genes that might explain why the retina looks different in the first place. That's the difference between a tool that spots a pattern and one that starts to explain it. It's the difference between correlation and the beginnings of a mechanism.

The judges at the 2026 Regeneron Science Talent Search clearly saw the same thing. RetinaMind took second place and a prize of $175,000, and Society for Science president Maya Ajmera singled out the project for combining computational sophistication with genuine biological depth.

None of this means a retinal scan will be replacing a clinical diagnosis any time soon, and Kang is refreshingly upfront about that. This is a proof of concept, built by one student on public datasets, and it has years of validation, regulatory scrutiny and clinical testing ahead of it before it belongs anywhere near a doctor's office. But it's worth sitting with what it points towards. Getting a diagnosis for autism or ADHD today can take months, sometimes years, of behavioural assessment, waiting lists and second opinions, not to mention the cost. A tool that could one day give families and clinicians an earlier, less invasive signal to work with, alongside everything else they already rely on, is a genuinely humane use of the technology. Kang himself has said that the conversations that mattered most to him were the ones with families living with an autism or ADHD diagnosis, and that talking with them shifted how he thought about the project, from detection towards what a diagnosis should actually make possible for the people receiving it.

That last point is the one I keep coming back to, because it's the same instinct behind everything I'm trying to build at Paadia. The value of an AI system was never really about how clever the model is. It's about who the people building it are trying to serve, and whether the ones who've historically waited longest for good tools, families navigating a diagnosis, smaller organisations without a research budget, communities that get left out of the pilot programme, are actually the ones it reaches first.

Edward Kang isn't a regulator or a policy team. He's one teenager with a question that wouldn't leave him alone. But his project is a useful reminder of what "AI as a force for good" actually looks like in practice. Not a bigger model or a bolder claim, but curiosity paired with humility, a black box that opens itself up to scrutiny, and a builder who spent as much time listening to families as he did training his network.

If the future of AI is going to be one we can live with, and one we can be proud of, it will look a lot more like this. Careful, transparent, and answerable to the people it's meant to help. That's the story worth telling, and the one worth building towards.