Artificial intelligence is ascendant in medicine—from AI eye doctors to chatbot therapists. As medical databases balloon in size and complexity, researchers are teaching computers to sift through and identify patterns, hinting at a future in which machine learning algorithms diagnose disease all on their own.

Sometimes, algorithms pick up on early signs of disease that humans wouldn't even know to look for. Last week, researchers at the University of North Carolina and Washington University reported an AI that can identify autistic infants long before they present behavioral symptoms. It's a thrilling opportunity: Early detection gives autism neuroscience a big leg up, as researchers try to understand what goes wrong during development. But now clinicians and researchers have to figure out what they’ll do with that information—is it just a research tool, or will they one day begin diagnosing and treating autism before symptoms start? Especially when it comes to infants, it won't be easy to entrust medical care to a computer-generated guess.

In this study, researchers scanned the brains of 59 6 month-olds whose older siblings were already diagnosed with autism. By age two, 11 of those infants had received a diagnosis of autism. By training a machine learning algorithm on their behavior and earlier MRI data, the scientists built a model that predicted 9 of those 11 autism cases, with no false positives. The AI predicted autism around a year before the earliest age—around 14 months—that clinicians diagnose it based on behavior.

Most immediately, this model can help researchers understand how the disorder develops and find early interventions for autism. Right now, researchers tracking autism development focus on infant siblings of people with autism; they have 1 in 5 chance of developing autism, compared to around 1 in 100 for the general population. With machine learning, though, researchers could scan infant siblings and focus their research on those at the very highest risk, buying more statistical power.

The trouble starts when you try to apply those predictions to treatment, not just research. Once a machine can detect autism earlier than doctors, a whole new set of practical and ethical problems emerge.

In order for a predictive tool to be useful for parents and patients, it would have to be accurate and relatively universal. That's certainly not true of this new model, its creators acknowledge. It will selectively predict certain kinds of autism: those that can be diagnosed by age 2 (some kids can't be positively diagnosed until much later), and types of autism that tend to run within families. For a predictive model to be useful for the general population, researchers would need to train their algorithm on a much, much bigger group. They could also improve accuracy by layering on other emerging predictive algorithms—ones based on genes, eye movements, and even baby babbling.

It would also need to be accessible to the right people. “Obviously we don't think that every kid in the population can get a scan at 6 months of age,” says Joseph Piven, the senior author on the UNC study—MRI is just too expensive and time-consuming. But genetic tests and family histories could help pediatricians hone in on high-risk infants and offer a scan to them.

If you figured out accuracy and distribution, then you'd need to provide effective treatment for those early-stage identifications. “I think we really do have to be thinking about these advanced computational methods for detecting autism—and what we're going to do once we detect,” says Zachary Warren, a clinical psychologist at Vanderbilt University who reviews autism therapies for the Agency for Healthcare Research and Quality. That doesn't mean diagnosis—at least not until the Diagnostic and Statistical Manual of Mental Disorders defines autism by something other than its behavioral markers. Tom Insel spent 13 years at the National Institutes of Mental Health trying to develop exactly that kind of quantitative framework for psychiatry—based on genetics, behavioral data, and physiological cues—and failed by his own account.

So this new information is problematic to use: How can clinicians create an intervention for an infant who might develop autism? All of the researchers interviewed for this story agree that early detection and intervention for autism is better. But current autism therapies for babies and toddlers focus on their specific behavioral deficits—teaching children to communicate needs, to play with toys, and to have positive interactions with caregivers. How do you design a treatment when you don't know what those specific deficits will be?

“We know that symptoms for one child are so different from symptoms for another child, so we have to be careful about any blanket treatment that is just going to be applied without knowing what the individual's particular difficulties are going to be,” says Somer Bishop, a clinical psychologist at University of California, San Francisco. Any pre-symptom treatment would have to be effective for the lowest common denominator—mostly likely limiting interventions to behavioral therapies that could help a child regardless of his or her neurodiversity.

That's where algorithmic detection could be doubly useful. The UNC group’s next goal is to predict specific autism symptoms, correlating brain scans with future language difficulties, sensory sensitivities, social difficulties, or repetitive behaviors. “Our model is flexibly able to capture this complex pattern in the brain that sets up the foundation for behaviors,” says Robert Emerson, the primary author. And if you can predict symptoms, you can get a lot closer to identifying targeted disease pathways—and targeted preventive treatments, either behavioral or pharmaceutical.

But that's not likely to happen soon. "Oftentimes autism detection science vastly outpaces intervention science," says Warren. Which puts caregivers who think their child might be at high risk of autism in a pickle.

After this study came out, Piven says many parents of children with autism contacted him requesting scans of younger siblings. “Without this kind of concrete information, what parents hear from their pediatricians is ‘Well, let's just wait and see.’ Parents are rightfully worried,” he says. But with a predictive scan in hand, experimental early intervention on a child without symptoms could be a source of stress. “What's that going to do to parents in terms of their mental health, their ability to attend to other children, their spouse, and family matters and their job?” asks Bishop. “That's where I get worried about people panicking and running to try to seek out very intensive intervention.”

To Bishop, a promising middle ground would be to encourage parents to focus on strategies that could help any child, regardless of whether they end up developing autism or not. “There are things that you can do in the context of your everyday routine, at bath time and during diaper changing and feeding, to encourage your child to communicate with you and to play appropriately with you,” she says. As with machine learning, there’s no harm in more training data for babies to learn about the world, regardless of their challenges.