AI Vision Models May Not Think Like Primate Brains After All

They can mimic brain activity. Yet they may not be thinking the same way. York University scientists have pulled back the curtain on a disconnect between artificial vision systems and the primate brain — and the result is quietly unsettling for anyone who treats today’s AI models as faithful brain surrogates.

The team, led by Assistant Professor Kohitij Kar, flipped the usual test used to compare neural networks and brains. Instead of asking whether a model can predict brain responses to images, they asked the opposite: can recorded brain activity predict a model’s internal representations? The answer was a revealing no.

To probe the question, researchers exposed the models and primate neurons to a broad visual diet: 1,320 natural photographs and realistic synthetic images — animals, faces, vehicles and everyday objects placed in indoor and outdoor scenes — plus 300 altered versions such as outlines, drawings and stylized renditions. The varied set ensured the inspection wasn’t fooled by one narrow stimulus type.

On the surface, many modern artificial neural networks (ANNs) perform admirably. They forecast the firing patterns of recorded neurons in parts of the visual cortex that help us recognize objects. But performance in one direction masked a deeper mismatch. When the team used brain signals to predict the models’ internal features, the fit fell apart for numerous model components. Curiously, neural recordings from one brain could predict another brain’s neurons reasonably well, which highlights the asymmetry.

Kar and his colleagues interpret this asymmetry as evidence that ANNs often converge on correct answers through strategies the biological visual system does not appear to use. In plain terms: two systems that arrive at the same behavioral output might still be using different internal tricks. "Models that line up with the brain tend to better predict human behavior," Kar notes, "but parts of these models seem to rely on shortcuts absent from primate vision."

That matters beyond academic curiosity. Scientists increasingly use AI as a scaffold to study perception and to design experiments for clinical populations. If a model’s internal logic diverges from the brain’s, researchers risk mistaking model-specific strategies for genuine neural mechanisms. Sabine Muzellec, a postdoctoral fellow on the study, stresses the stakes: the field needs trustworthy diagnostics to separate genuinely brain-aligned components from those that only masquerade as such.

The team offers a practical tool: a reverse predictivity test that diagnoses which parts of a network truly reflect neural activity. That diagnostic could steer the next generation of models toward architectures and training regimes that capture not just external behavior but the inner computations that generate it. Think of it as choosing a faithful translator rather than a clever mimic.

There are broader implications. AI systems influence work in hearing, language and motor control as well as vision. If hidden mismatches exist in one sensory domain, they might lurk in others. For clinical research — autism studies, PTSD investigations, rehabilitation design — models that align more faithfully with neurotypical brain function could sharpen hypotheses and improve interventions. Conversely, overreliance on misaligned models risks misleading conclusions.

The message from York’s lab is both a caution and an invitation: treat current AI as powerful but imperfect mirrors of the brain. We can exploit their predictive power. We should also demand that the reflections they offer match more than just the surface.