AI Models Predict Brain Activity, But Brain Cannot Predict AI’s Internal Processes, Study Finds

Artificial intelligence models commonly used to study human vision may operate through internal processes that differ significantly from those of the human brain. While these systems can predict neural activity, a new study reveals that the brain’s own signals cannot reliably predict how the models themselves function internally.

Testing visual processing limits

For over a decade, researchers have utilized artificial neural networks to gain insight into how the human visual system identifies objects. These systems are often characterized as “brain-like” because of their capacity to mirror patterns of brain activity, ScienceDaily reported. “Until now, scientists mostly tested this in one direction. They asked whether AI models can predict brain activity,” York University Assistant Professor Kohitij Kar said.

The researchers analyzed how these systems processed 1,320 natural photographs and realistic synthetic images, according to ScienceDaily. These images depicted 10 familiar subjects, including a bear, an elephant, a face, an apple, a car, a dog, a chair, a plane, a bird, and a zebra. The team further introduced 300 images showing these subjects in various styles, such as drawings, outlines, and artistic interpretations, dailyartifact.ai noted.

Diverging internal mechanisms

The study found a clear disconnect when the direction of the analysis was reversed. While the AI models accurately predicted recorded neuronal responses, the biological brain’s activity did not effectively predict the models‘ internal features, ScienceDaily reported. This disparity suggests that the models achieve similar output results to the brain while relying on entirely different internal strategies.

“The results were striking. While AI models can predict the neurons we recorded in the brain fairly well, the brain cannot equally predict many of the model’s internal features,” said Kohitij Kar, an assistant professor at York University. Kar said that this lack of alignment does not occur when comparing neuronal activity between two different human brains.

AI may not work like the human brain after all, scientists discover
Photo: dailyartifact.ai

Risks for clinical research

This discovery matters for scientists who use AI to model human behavior, language, or neurological conditions. If researchers rely on models that appear accurate on the surface but function differently underneath, they risk drawing misleading conclusions about human neural mechanisms, themorning.science reported. The findings suggest that current AI systems may incorporate strategies that the human brain does not use, according to Newest Science Information.

“Our findings challenge how similar current AI systems really are with the primate brain,” said Sabine Muzellec, a postdoctoral fellow at York University, as told to ScienceDaily. “We show that models that were previously thought to be brain-like rely on internal components that the brain does not appear to use.”

Refining AI for Scientific Reliability

Rather than suggesting that researchers abandon artificial neural networks, the study provides a pathway to make these tools more scientifically reliable. According to Newest Science Information, the findings offer a method to improve how developers build AI that more closely matches the human brain.

Story on The Morning Science
Photo: themorning.science

Neuroscientists who rely on these models to study biological processes may need to reconsider how much their simulations actually reflect the brain. By utilizing the new diagnostic framework, researchers can better identify which internal components of a model truly align with biological mechanisms, as reported by ScienceDaily on November 21, 2026.

Kohitij Kar is an assistant professor in the Faculty of Science at York University, where his research focuses on the intersection of neuroscience and artificial intelligence. As a member of the York-led Connected Minds initiative, Kar investigates how biological vision processes information to better understand the computations underlying human perception and behavior.

His work emphasizes that to develop AI that is human-compatible, researchers must look beyond output accuracy and instead evaluate whether the internal mechanisms of these models align with those of the human brain.

Improving future models

The researchers do not suggest abandoning these tools, but rather using their new diagnostic metric to refine them. By identifying which parts of a neural network truly align with biological processes, developers can build more reliable models for investigating conditions such as autism or post-traumatic stress disorder, ScienceDaily reported.

The team has released a publicly available testing toolkit for developers to evaluate their own models. The full study, which offers a new framework for evaluating artificial intelligence beyond simple task performance, was published in Nature Machine Intelligence.

Photo of author

Technology Desk

Technology Desk at Archyde.

Haas Retains Câmara for Final Two 2026 Races, Ending F2 Title Bid