Scientists at the University of York have discovered a significant flaw in Artificial Intelligence models, challenging the widespread belief that machine algorithms operate similarly to the human brain. Although artificial neural networks have long been considered the best models for the human cognitive system, the new research shows that their internal strategies do not fully align with the principles of primate brain function.
AI models are increasingly being used in clinical trials to study human behavior, including treatments for autism and Post-Traumatic Stress Disorder (PTSD). If a model perceives the world differently from a human, its use in experiments could be hazardous.
What does the study say?
Until now, scientists tested AI only one-dimensionally, observing how effectively a program could predict human brain reactions. Researchers at the University of York, however, logically reversed this approach.
Their theory is based on the premise that if AI is truly a digital copy of the human brain, the connection must be bidirectional. In other words, by observing our brain activity, we should be able to precisely determine which operations are being performed within the AI model.
To test this theory, researchers conducted a large-scale experiment using over 1,600 different types of images, ranging from real photos of animals to schematic drawings and artistic variations.
The study revealed that while AI can explain the behavior of human neurons relatively well, the human brain cannot predict AI’s internal computations with the same accuracy. This suggests that the similarity between them is only superficial, and the programs operate on entirely different principles.
Ultimately, the research confirmed that modern AI processes visual information using strategies that nature and human evolution do not utilize at all. This discovery is critically important; if AI perceives the world differently than we do, its application in medical research or psychological experiments could lead to erroneous conclusions. Scientists are now working to align these programs more closely with the actual biological principles of brain function.
Researchers have already developed specialized testing tools to help AI developers improve their models. The goal is to create Artificial Intelligence that is maximally compatible with the human brain. This is particularly vital for modeling neurotypical brains, which serves as the foundation for future medical research.
