Autism: Capturing Clinical Certainty with AI

With artificial intelligence, a Montreal-based research team is seeking to detect and refine the certainty that guides autism diagnosis, a certainty now resting more on repetitive behaviours and sensory responses than on the social criteria that have been at the core of the DSM-5 for forty years. 

A child under eighteen months learns the letters of the alphabet without being taught, in a language his parents do not speak, from a tablet. He recognizes their shapes, their curves, their recurrence, long before grasping their meaning. This is an autistic signal still largely absent from public discourse, unlike hand-flapping, more familiar, but no less a sign that a clinician identifies instinctively.

This is the kind of certainty that researcher Danilo Bzdok wants to capture.

A professor at McGill University and a researcher with IVADO’s AI and Neuroscience research group, Danilo Bzdok starts from an unsettling observation: after decades and billions invested in genetics and brain imaging, no technology surpasses the precision of a reasonably trained psychiatrist talking to a child. Thousands of studies have attempted to improve autism diagnosis from the outside. None has outperformed clinical expertise.

Decoding what clinicians know 

This is the context in which Danilo Bzdok partnered with Laurent Mottron, professor in the Department of Psychiatry and Addiction Medicine at Université de Montréal and one of the world’s leading authorities on autism. The full team includes Emmett Rabot, Siva Reddy and Eugene Belilovsky. 

Without being an autism specialist himself, Danilo Bzdok is interested in what data can reveal where conventional technologies have failed. Laurent Mottron, for his part, has made more than 3,000 diagnoses over his career. He brings what AI cannot generate on its own: the clinical filter that precedes the question.  

The study, published in March 2025 in Cell, submitted more than 4,200 clinical reports written by healthcare professionals with between 5 and 20 years of experience to analysis by large language models, AI systems capable of identifying regularities across thousands of texts and detecting which terms appear significantly more often in reports of children diagnosed as autistic. These reports had been written freely, without instruction to follow the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders). The clinicians described what they observed, in the course of working toward a diagnosis they had not yet made at the time of writing. 

The DSM-5 places social deficits at the heart of autism diagnosis: emotional reciprocity, non-verbal communication, the quality of relationships. Evaluations have been built around these dimensions for forty years. Danilo Bzdok’s model found the opposite in clinicians’ notes. What they most often described for children diagnosed as autistic was not social withdrawal. It was repetitive behaviors, specific interests and sensory responses. The term “flapping,” that characteristic hand movement, appears 21.5 times more often in reports of confirmed autistic children than in those where the diagnosis was ruled out. 

“The degree of importance is inverted,” says Danilo Bzdok. “We’re not saying the social factor is irrelevant. We’re saying the relative weight is the opposite of what the field has believed for forty years.” 

Diagnostic domains: a reversed priority 

Social-communicative domain: social and emotional reciprocity, non-verbal communication, developing and maintaining relationships.

Domain of repetitive behaviors and restricted interests: repetitive movements, intense and restricted interests, sensory sensitivity, resistance to change.

Observed behaviors: concrete examples

Source : Jack Stanley, Emmett Rabot, Siva Reddy, Eugene Belilovsky, Laurent Mottron et Danilo Bzdok. Cell, mars 2025. 

Laurent Mottron has an evocative phrase for what AI made possible: “Doing hard science with soft material. You take informal descriptions and they become a better predictor than genetics and imaging combined.”

This knowledge has a name in Laurent Mottron’s practice: clinical certainty. It can be graded: a clinician can be asked how confident he is in his diagnosis, and that degree has concrete consequences. A highly confident clinician moves forward; a moderately confident one sees the child again, consults a colleague. This gradable, operational certainty is what Danilo Bzdok’s study seeks to make accessible. 

Tools without specificity 

Even when experienced clinicians outperform the DSM’s official criteria, the rest of the system is not yet ready

“The current problem is overdiagnosis,” says Laurent Mottron flatly. An industry has grown up around diagnosis: private evaluations costing two to three thousand dollars, conducted with tools that, in his view, “have no specificity whatsoever. With those tools, virtually every poorly defined psychiatric condition comes back positive.” 

A child with severe ADHD is the clearest example: his impulsive behavior eventually drives other children away, until his social life collapses. On standard evaluations, which measure the outcome rather than the cause, this isolation resembles that of an autistic child, without the true, and different, cause ever being investigated. 

An autistic child does not avoid the group because he was excluded from it; he withdraws because contact, noise and unpredictability cost him more than they give him.

This concern is not an isolated one. In a British scientific journal this year, researcher Uta Frith, one of the founding figures behind the concept of the autism spectrum, questions whether the clinical meaning of autism has become diluted. 

Identifying the real signals 

This is where the combined work of Bzdok and Mottron offers something unprecedented: concrete markers for families, drawn from thousands of cases. 

Hand-flapping comes first among these signs, but not just any kind. Laurent Mottron’s team submitted hundreds of videos to experts around the world, blinded, as part of research whose results are in the process of being published. All of them recognized certain movements as autistic with a high degree of agreement across continents. 

“Everyone flaps their hands,” the researcher explains,“ but experts, completely independently, recognize one type of flapping as more autistic than others.” 

A second signal ties back to the opening of this article: the child’s marked interest in the alphabet, drawn from recent work by Laurent Mottron’s team. The draw is not social. It is directed toward shapes and patterns themselves. 

“This may be the strongest signal to emerge from our recent work,” says Laurent Mottron. “It explains both how autistic people learn language and how they do not.” 

Recent research from his team, involving 700 subjects, confirms this sign, which is absent from the DSM. 

Hands over the ears as well, but only in a persistent and specific form: the gesture that recurs several times a day for years in front of a flushing toilet, the gear changes of trucks, or certain whistling sounds. 

Not all signs of sensory sensitivity carry the same diagnostic weight. The clothing tag irritating the back of the neck has circulated widely as a sign of autism. Laurent Mottron is direct: it is not one. These manifestations are too common in the general population to constitute a reliable diagnostic marker. 

And then there is familial risk: when a child is diagnosed as autistic, the probability that the next child in the same family will also receive a diagnosis is around 25 per cent. “That is the most solid finding in the neuroscience of autism,” says Laurent Mottron. 

Steps still to take 

When Danilo Bzdok is asked what a clinician might now change in their practice, his answer is immediate: “They’re not allowed to. They must continue following the official criteria.” 

The longer-term goal: for diagnostic criteria to be revised not by expert committees debating among themselves, but by the data itself. 

“We may be in the process of rethinking how a diagnostic category gets created,” says Laurent Mottron. 

This experienced clinician who instinctively recognizes the repetition of a gesture or a child’s pull toward a shape cannot yet change his practice. But what he carried as a solitary certainty, large language models have now shown to run through more than four thousand clinical reports, written the same way, by professionals who never consulted one another.