Artificial Intelligence Models Differ From Brains, York Study Finds
York University researchers found that AI models can predict brain-cell responses to images without processing visual information in the same way as the brain. The team has released a testing toolkit to help scientists identify which parts of their models correspond to neural activity.

York University researchers have found that artificial intelligence models can predict brain-cell responses to images without necessarily processing them like the brain. The study tested artificial neural networks, computer systems that can perform tasks such as recognizing objects in pictures.
The distinction matters for scientists who use those models to investigate how people see. A model that produces a brain-like response may still rely on different internal processes, the researchers found.
“Artificial intelligence systems are often described as 'brain-like' because they can predict activity in parts of the brain that help us recognize objects,” said York University Assistant Professor Kohitij Kar, the study’s senior author.
Researchers Tested Predictions in Reverse
The York University team tested whether recorded brain activity could predict features inside AI models. Scientists had generally tested the relationship in the other direction: whether models could predict brain activity, Kar said.
The researchers developed a method called reverse predictivity testing. It examines whether similarities between models and brain activity hold when the prediction runs both ways.
AI models generally predicted the responses of recorded neurons, the specialized cells that transmit information in the brain. The neurons’ responses were much less effective at predicting many internal features of those models.
Kar said the same problem did not appear when researchers compared neural activity across different brains. The result suggests that a model can match brain responses while using different internal mechanisms.
“The findings suggest that today's AI systems solve visual tasks partly using internal strategies that the brain may not use,” Kar said. He added that the parts of AI models aligned with the brain were better at predicting human behavior.
Images Tested Across Visual Styles
The team examined responses to 1,320 natural photographs and realistic synthetic images. Those pictures showed 10 subjects: a bear, an elephant, a face, an apple, a car, a dog, a chair, a plane, a bird and a zebra.
The subjects appeared against natural scenes and indoor or outdoor backgrounds. Another 300 images showed the same subjects as outlines, drawings, simplified representations and artistic interpretations.
York Postdoctoral Fellow Sabine Muzellec, who worked on the study, said models previously considered brain-like relied on internal components the brain did not appear to use.
Team Releases Testing Toolkit
The York University researchers released a public toolkit for developers to test their AI models. The approach helps identify which model features correspond to brain activity, Kar said.
Kar said the method was especially important for the team’s autism research program. That work uses models of the neurotypical brain, meaning one with neurological development and functioning generally considered typical, as a baseline.
The study was published in Nature Machine Intelligence.
Take the quizWhat Kind of Holiday Spender Are You? Take the Black Friday QuizStart →
Markets · FintechHow to Calculate Compound Interest: Formula, Examples and the Rule of 72
Lifestyle · Food & HealthNormal Blood Pressure Chart by Age: Healthy Ranges Explained
Lifestyle · Food & HealthWhat Is Raynaud’s Syndrome? Why Fingers Turn White in the Cold
Lifestyle · Food & Health4 Ways Walking After Meals Helps Your Body, Doctors Say

















