“Artificial intelligence systems are often described as ‘brain-like’ because they can predict activity in parts of the brain that help us recognize objects,” Kohitij Kar told ScienceDaily.
Artificial intelligence models widely used to study human vision may operate fundamentally differently from the biological architecture of the human brain, according to a study published on November 21, 2026. Researchers at York University found that artificial neural networks can accurately forecast brain activity while relying on entirely distinct internal mechanisms, challenging long-held assumptions in computational neuroscience.
For over a decade, scientists have relied on ANNs as leading tools to investigate how vision works, treating high predictive power as proof of mechanistic alignment between synthetic weights and biological synapses. The York University team discovered a significant disconnect by approaching the problem in reverse. While artificial models successfully predict recorded neural responses, biological brain activity does not reliably predict internal features within the models.
The research team examined responses to over 1,600 images, including natural photographs and realistic synthetic images featuring 10 familiar subjects, alongside 300 additional images in varying artistic styles. They applied a reverse predictivity testing method to evaluate similarities that conventional tests overlook. Sabine Muzellec, a postdoctoral fellow and trainee with the university’s Connected Minds program, said that previous models relied on components the brain does not appear to use, according to ScienceDaily.
The findings carry major implications for researchers who use artificial intelligence to design experiments, study human behavior, and investigate neurological conditions. Conclusions drawn from AI models about human vision or behavior could be misleading if the underlying internal workings differ from biological processes.
Computational neuroscientists frequently treat artificial networks as working hypotheses for how biological neural circuits solve complex sensory and cognitive tasks. Relying purely on surface-level predictive alignment risks creating an illusion of biological plausibility, as modern deep learning models function as complex black boxes that can mask mathematical artifacts.
Kar warned that ignoring these underlying differences could lead scientists down misleading avenues as models continue to advance. Developing AI systems that more faithfully reproduce brain function could eventually improve research into conditions ranging from post-traumatic stress disorder to autism.
Rather than suggesting researchers abandon artificial neural networks, the study provides a diagnostic metric to make them more scientifically reliable. The team released a publicly available testing toolkit that developers can use to evaluate their models and identify opportunities for improvement.
“Our approach helps identify which parts of an ANN truly match brain activity, allowing us to build more reliable models for understanding how people see and interpret the world,” Kohitij Kar, senior author of the study, told ScienceDaily.
The York University team developed a reverse predictivity testing method to check whether recorded brain activity could predict what was happening inside the AI models.
The study was reported by ScienceDaily on November 21, 2026.
Established in 1959, York University began its first classes in September 1960 with a total of 76 students before eventually growing into Canada’s third-largest university. Over the decades, the institution achieved several national milestones, including launching the country’s first undergraduate program in dance in 1970 and its first PhD program in women’s studies in 1992.
“We provide a well-vetted diagnostic metric for the field,” Sabine Muzellec told ScienceDaily.