Neuroscience & AI

AI Models May Not Think Like the Brain After All — A Reverse Test Reveals the Gap

Artificial neural networks have become the leading tools for explaining how the brain processes sight. But when researchers at York University turned the test around — asking whether brain activity could predict what happens inside the AI — they found a striking asymmetry.

• York University researchers tested artificial vision models with 1,320 natural photographs and 300 additional altered images. While AI models could predict recorded brain activity fairly well, the brain could not equally predict many of the model's internal features — a mismatch that suggests AI reaches correct answers through processes the brain does not use.
• The study, led by Canada Research Chair in Visual Neuroscience Kohitij Kar and postdoctoral fellow Sabine Muzellec, introduces a "reverse predictivity" test as a diagnostic tool for the field. When neurons from one brain were compared against neurons from another brain, the prediction worked both ways — but the AI-brain comparison did not.
• The findings challenge the widespread assumption that today's artificial neural networks process visual information in ways that resemble the primate brain. As AI models become more complex, the mismatch could widen unless researchers actively address it, which has implications for neuroscience research and clinical studies that rely on AI models to understand human behaviour.

Over the past decade, artificial neural networks (ANNs) — computer models designed for visual tasks — have become some of the leading tools for explaining how the brain processes sight. These systems are often described as "brain-like" because they can predict activity in parts of the brain that help humans recognise objects. But researchers at York University wanted to determine whether these systems truly operate like biological vision, or whether the resemblance is only surface-deep.

"Until now, scientists mostly tested this in one direction. They asked whether AI models can predict brain activity," says York University Assistant Professor Kohitij Kar, senior author of the study. The researchers reversed that familiar test. If AI genuinely reflects the brain, they reasoned, then recorded brain activity should also predict the model's internal responses. To examine this possibility, they developed a reverse predictivity test.

The team tested the models with 1,320 natural photographs showing bears, elephants, faces, apples, cars, dogs, chairs, planes, birds, and zebras against indoor, outdoor, and other natural backgrounds. They also used 300 additional images depicting the same objects in altered forms — outlines, drawings, simplified representations, and artistic variations — to test whether the relationship between brain activity and AI features held across different visual styles.

The results were striking. While AI models could predict the neurons recorded in the brain fairly well, the brain could not equally predict many of the model's internal features. Importantly, this asymmetry disappeared when comparing one brain against another — neurons from one brain could predict neurons from another brain in both directions. The imbalance suggests that ANNs may reach correct visual answers through processes that differ from those used by primate brains.

Kar warns that the mismatch could widen as models become more complex unless researchers address it early. "The findings suggest that today's AI systems solve visual tasks partly using internal strategies that the brain may not use," he says. Importantly, the parts of AI models that do align with the brain are also better at predicting real human behaviour, suggesting that improving this alignment could lead to both better AI and better neuroscience.

Researchers increasingly use AI models to design studies of human behaviour, including clinical research. Much of that work assumes the systems process the world in ways that resemble the human brain. "Our findings challenge how similar current AI systems really are with the primate brain," says Muzellec. "We show that models that were previously thought to be brain-like rely on internal components that the brain does not appear to use. We provide a well-vetted diagnostic metric for the field."