A new reverse-prediction study suggests that AI systems often described as “brain-like” may actually rely on visual strategies that primate brains do not use. Although modern artificial neural networks can sometimes predict patterns of brain activity during object recognition, that similarity may not reflect the same underlying computations.

Researchers at York University investigated whether leading computer vision models truly process visual information in a way that resembles biological vision. According to senior author Kohitij Kar, previous research largely tested the relationship in only one direction—asking whether AI models could predict neural activity in brain regions involved in recognizing objects.

Figure 1. AI and Primate Brains Process Visual Information Differently

The new study challenges the assumption that successful prediction automatically means AI and the brain use the same mechanisms to solve visual tasks. Figure 1 shows AI and Primate Brains Process Visual Information Differently.

The researchers reversed the traditional AI–brain comparison. They reasoned that if AI models truly mirror the brain’s visual processing, then recorded brain activity should be able to predict the models’ internal responses, not just the other way around. To test this idea, they created a reverse predictivity test that examined whether neural activity from primates could accurately forecast the hidden activity patterns inside AI vision models.

Kohitij Kar explained that the ultimate goal is to develop computational models that reveal the actual neural mechanisms underlying object recognition and visual perception. Humans effortlessly recognize objects and track their movement in everyday life, but reproducing that capability computationally remains an exceptionally difficult scientific challenge.

Reverse Prediction Test Challenges Brain-Like AI

The research team, including York postdoctoral fellow Sabine Muzellec, evaluated the AI models using 1,320 natural and realistic synthetic images. The dataset contained a wide variety of objects—including bears, elephants, faces, apples, cars, dogs, chairs, airplanes, birds, and zebras—presented in indoor, outdoor, and other natural scenes.

To further challenge the models, the researchers added 300 transformed versions of the same objects, such as outlines, sketches, simplified depictions, and artistic renderings. Using this diverse image set allowed them to test whether the connection between brain activity and AI representations remained consistent across very different visual styles rather than only standard photographs.

Brain Activity Reveals a Hidden AI Mismatch

The reverse prediction test revealed a striking asymmetry. Modern AI vision models were able to predict recorded brain activity reasonably well, but brain activity could not predict many of the models’ internal features to the same extent. In contrast, neural activity from one brain could successfully predict activity in another brain, suggesting that biological brains share a common representational structure that many AI models do not.

This imbalance indicates that artificial neural networks may arrive at correct visual judgments using computational strategies that differ from those used by primate brains. As models become more complex, that mismatch could become even larger unless it is addressed directly. An AI system that predicts neural responses but contains internal representations that cannot be recovered from brain activity may therefore be a poor explanation of how biological vision actually works.

The findings matter because AI models are increasingly used to guide research on human perception, behavior, and neurological disorders. Much of this work assumes that the models process information in a brain-like manner, but the new results suggest that this assumption may be weaker than previously thought.

According to the researchers, many AI systems that appear brain-like may depend on internal components that primate brains do not use. To address this, the study introduces a reverse predictivity metric that can help identify which parts of an artificial neural network genuinely correspond to biological brain activity.

Improving this alignment could make AI-based research more reliable in areas such as autism, PTSD, language, hearing, and movement disorders [1]. The researchers argue that poorly aligned models could lead to misleading interpretations of human behavior, whereas models whose internal representations better match neural activity may provide a stronger foundation for studying how people perceive and interpret the world.

Kar notes that this is particularly important for autism research, where models of the neurotypical brain are often used as a baseline for understanding differences in perception and cognition.

The study also found that the portions of AI models that most closely align with brain activity are the same components that best predict real human behavior, suggesting that improving brain alignment may also improve the human relevance of AI systems.

Reference:

1. https://scitechdaily.com/scientists-discover-ai-models-may-not-think-like-the-brain-after-all/

Cite this article:

Janani R (2026), Scientists Find AI Models May Not Resemble the Brain After All, AnaTechMaz, pp. 1025