The notion that artificial intelligence mirrors biological functioning has long been a cornerstone of AI development. However, new research from York University suggests this resemblance is largely superficial. While artificial neural networks (ANNs) can accurately predict brain activity during object recognition, the reverse is not true: biological neural activity cannot predict the internal responses of AI models.
The Reverse Predictivity Test
To expose this asymmetry, a team led by Assistant Professor Kohitij Kar developed a reverse predictivity test. While previous studies focused on AI's ability to simulate the brain, this study examined whether primate neural activity could explain a model's internal processes. Using 1,320 natural photographs and 300 stylized images, the results revealed a striking mismatch.
While AI models mimic neurons effectively, the brain does not reflect the model's internal features. This suggests that AI reaches correct visual answers through computational paths entirely different from biological ones. Essentially, the AI is not "seeing" like a primate but is employing mathematical shortcuts to achieve the same output.
Beyond the Biological Metaphor
This discovery aligns with a broader debate on synthetic reasoning. Recent research, including studies by Apple on Large Reasoning Models (LRMs), has shown a collapse in logic when facing classic puzzles, suggesting that AI does not "think" cognitively. Similarly, research published in PNAS Nexus indicated a total breakdown in focus for advanced AI models as cognitive demands increased.
The danger lies in using poorly aligned systems to interpret human behavior, particularly in clinical settings. Using AI models as baselines for autism or PTSD research requires rigorous validation to ensure that data interpretation isn't skewed by non-biological computational strategies.
Toward Efficient and Transparent AI
Despite the gap, achieving alignment between biology and silicon remains a strategic goal. If AI systems could more closely mimic brain structures, they might learn more efficiently from less data, significantly reducing energy and infrastructure demands.
To facilitate this progress, the York University researchers have released a testing toolkit allowing developers to evaluate how closely their model's internal features correspond with actual brain activity, moving the AI "black box" toward a more transparent and scientifically grounded tool.