AI models mimic brain activity without working like biological vision, study finds
York University researchers found that brain cells fail to predict the inner workings of vision AI, showing that machines use fundamentally different methods to recognize objects.
For years, neuroscientists have turned to artificial neural networks to study vision. Because these computer systems can predict how cells in the brain react when viewing an image, many researchers assumed they process visual data like biological brains. A new study from York University shows that this relationship only works one way.
The one-way mirror in vision models
Scientists usually test vision models by checking whether the software can forecast brain cell activity. The York University team decided to flip the process. They created a technique called reverse predictivityReverse predictivityA test measuring whether biological brain data can predict the internal features of a computer model. Open in glossary testing to see if recorded brain responses could forecast the internal states of the AI models.
The team tested this using 1,320 natural and synthetic pictures, along with 300 stylized images such as drawings and outlines. The images depicted 10 common items, including faces, cars, apples, dogs, and birds. While the AI successfully mirrored how biological neurons reacted to the images, the biological neurons struggled to forecast the internal processes of the AI. When comparing brain cells from one living subject to another, this disconnect did not happen.
Risks for medical and behavioral research
The findings indicate that an AI system can match biological results on the surface while using completely different inner mechanics. Researchers warn that continuing to build models using only one-way testing will cause these systems to diverge further from biological reality.
This gap carries consequences beyond computer vision. Scientists currently rely on these systems to model human behavior and study health conditions such as autism and post-traumatic stress disorder. Similar models also support research into language, movement, and hearing. If the internal workings of the software do not match biology, researchers risk reaching incorrect conclusions. The study authors suggest using two-way testing to isolate the specific parts of artificial networks that genuinely match real nervous systems.
How does this story land?
One tap per reader.
Quick poll
Should neuroscientists stop using standard AI models to study human perception?
One tap. No sign-up.
Ask this story
Still have a question?
Answers use only what is in this story. They are written by AI and can be wrong.