Science
Gist from Sciencedaily

AI Visual Processing May Diverge Significantly From Human Brain, York Study Finds

Summarized November 22, 2026
Jump to key takeaways

A Critical Gap in AI Brain Models

Researchers at York University have identified a fundamental flaw in how artificial neural networks are used to study human vision and brain function. For the past decade, these AI models have become central tools in neuroscience, valued for their ability to predict patterns of neural activity in the visual cortex. The prevailing assumption has been that if an AI system can successfully replicate how the brain responds to visual stimuli, it must be processing information in a similar manner. A new study challenges this assumption directly, suggesting that AI systems may achieve brain-like outputs while using entirely different internal mechanisms.

Assistant Professor Kohitij Kar and his team approached the problem from an unconventional angle. Rather than simply testing whether AI could predict brain activity—the standard methodology—they flipped the equation and asked whether brain activity could equally well predict the internal workings of AI models. This reverse predictivity testing revealed a striking asymmetry that conventional one-directional tests had missed.

The Experiment and Surprising Results

The research team examined how both artificial neural networks and actual primate brains responded to 1,320 natural photographs and synthetic images featuring ten familiar objects: bears, elephants, faces, apples, cars, dogs, chairs, planes, birds, and zebras. They also tested 300 additional images showing the same subjects rendered as outlines, drawings, simplified shapes, and artistic interpretations. This comprehensive visual stimulus set allowed researchers to thoroughly probe how visual systems process diverse representations of the world.

The results revealed a profound disconnect. While AI models successfully predicted the responses of recorded neurons—demonstrating that they produce outputs resembling brain activity—the reverse relationship failed to hold. When researchers used actual neural recordings to predict the internal features of AI models, the predictions were far less accurate. Crucially, this bidirectional problem did not appear when comparing neuronal activity from one animal brain to another, suggesting it reflects a fundamental difference between artificial and biological vision processing rather than general measurement noise.

Kar emphasized the significance of this finding: artificial systems can generate results that closely resemble those of the brain while relying on very different computational strategies. A system might achieve correct answers through completely different reasoning paths—producing the right output for the wrong reasons.

Implications for Neuroscience and Beyond

The implications extend far beyond academic curiosity about AI architecture. Neuroscientists increasingly rely on artificial neural networks to design experiments, interpret brain imaging data, study human behavior, and investigate conditions ranging from autism to post-traumatic stress disorder. If these models operate through fundamentally different mechanisms than actual brains, researchers could reach misleading conclusions based on systems that appear accurate on the surface but function differently underneath.

The potential problems multiply across neuroscience domains. Similar AI models are already deployed to study hearing, language processing, and motor control. If the visual processing disconnect extends to these other sensory and cognitive systems, entire research programs may rest on flawed assumptions about how biological nervous systems work.

Kar warned that this problem could worsen over time. If researchers continue refining AI models based solely on their ability to predict neural activity, without checking whether their internal mechanisms align with actual brain function, the gap between models and biology could accumulate. Each iterative improvement toward better neural predictions might actually push models further from genuine biological mechanisms.

A Path Forward

Rather than abandoning artificial neural networks in neuroscience, the findings point toward more rigorous development practices. By using reverse predictivity testing to identify which AI components genuinely correspond to real neural activity, researchers can build models that more faithfully reflect biological mechanisms. This matters especially for studies of conditions like autism, where understanding how typical brains process information provides essential baselines for comparison.

The research team has released a publicly available testing toolkit so other scientists can evaluate their own AI models using the reverse predictivity method. This democratization of the diagnostic tool could help the field systematically improve the biological authenticity of its computational models.

The study, published in Nature Machine Intelligence, argues for a fundamental shift in how researchers evaluate AI systems designed to explain brain function. Performance on tasks matters, but equally important is whether the system reaches correct answers through processes that genuinely resemble those of the human brain. As artificial intelligence becomes increasingly embedded in neuroscience research, distinguishing between superficial similarity and genuine correspondence will shape the reliability of discoveries about human perception and brain function for years to come.

Key Takeaways

  • AI predicts brain activity well, but brain cannot reciprocally predict AI internals
  • Models producing brain-like outputs may use fundamentally different internal mechanisms
  • Bidirectional prediction gap does not appear comparing brain to brain
  • Neuroscience research on autism, PTSD relies on potentially flawed AI assumptions
  • Reverse predictivity testing identifies AI features that genuinely match brain activity
  • Similar vision-hearing-language AI models may harbor same biological mechanism mismatches
  • York team released public toolkit for researchers to evaluate model biological authenticity
Read original article at Sciencedaily

Summarize any article in seconds

Gist is a free Chrome extension that instantly creates AI-powered summaries and key takeaways for any article or podcast on the web.

Add to Chrome — It's Free
⚡ Instant summaries 💬 Chat with articles 🔒 Privacy-first