Reverse predictivity testing is at the center of new research suggesting artificial intelligence may be less brain-like than long assumed. A York University team reports a key weakness in popular AI models used to probe how the brain processes visual information.
For more than a decade, artificial neural networks, or ANNs, have been used as stand-ins to examine vision, including object recognition. Their success at recreating patterns of neural activity has made them go-to tools in neuroscience.
The latest findings pose a basic challenge. If AI can approximate certain brain responses, does that mean it handles visual information the same way?
York University Assistant Professor Kohitij Kar, senior author of the study, said AI systems are often labeled brain-like because they predict activity in areas that support object recognition. Most tests have run in one direction, asking whether AI can forecast brain activity.
Reverse predictivity testing highlights a hidden gap
The researchers flipped the question. Instead of only checking how well AI predicts neuronal signals, they examined whether recorded brain activity can predict internal states of the AI models.
If the systems truly operate like the brain, the relationship should be bidirectional. To test this, the team developed reverse predictivity testing, a method designed to capture similarities that standard evaluations might miss.
Kar, the Canada Research Chair in Visual Neuroscience and a member of York’s Centre for Vision Research and Centre for Integrative and Applied Neuroscience, noted the broader goal is to build computational models that reveal mechanisms of object recognition and motion perception, tasks that are simple for humans yet computationally difficult.
The team, which included York Postdoctoral Fellow Sabine Muzellec, a trainee in the university’s Connected Minds program, measured responses to a diverse image set.
The test battery featured 1,320 natural photos and realistic synthetic images across 10 categories: bear, elephant, face, apple, car, dog, chair, plane, bird, and zebra, shown in varied natural scenes and indoor or outdoor contexts. They added 300 more images presenting the same subjects as outlines, drawings, simplified renderings, and artistic styles.
AI predicts brain signals, but the reverse falls short
The results point to a notable disconnect between AI and biological vision.
AI models generally predicted responses of recorded neurons. Yet when the process ran in reverse, those neuronal signals were far less effective at predicting many internal features of the models.
This contrast stood out because the team did not see a similar shortfall when comparing neuronal activity across different brains.
Kar said that while AI can forecast recorded neurons reasonably well, the brain cannot equally account for many internal components of the models. That pattern did not appear when comparing neurons from one brain with another.
The findings indicate an AI system can produce outputs that look brain-like while relying on different internal mechanisms. Matching responses does not guarantee matching processes.
Kar cautioned that refining models solely to improve predictions of neural activity could widen the gap with actual primate brains over time. That would undercut a central motivation for using such models, which is to build realistic explanations of brain function.
He added that model components most aligned with the brain also tend to better predict human behavior.
Implications for neuroscience and behavior research
The impact extends beyond vision. AI increasingly supports experiment design, behavioral studies, and research into neurological and psychological conditions.
Many uses rest on the premise that AI processes information in ways similar to the human brain.
If that premise is wrong, researchers risk drawing incomplete or misleading conclusions from models that appear accurate at the output level but differ internally.
Muzellec said the findings question how similar current AI systems are to the primate brain. According to her, models once considered brain-like may depend on internal elements the brain does not use, and the study provides a vetted diagnostic metric for the field.
Closer brain-matching AI could aid research into conditions ranging from post-traumatic stress disorder to autism. For now, the observed differences argue for caution when using existing models to study human behavior or to guide clinical work.
The concern may reach other domains in neuroscience, since similar AI models are used to examine hearing, language, and movement. If their internal processes also diverge from biology, researchers may need to reassess how much these tools can reveal about the nervous system.
Building models that better mirror brain mechanisms with reverse predictivity testing
The study does not argue for abandoning ANNs. Instead, it offers a way to make them more informative and reliable.
By pinpointing which model features correspond to real neural activity, researchers can concentrate on systems that more faithfully reflect biological mechanisms of perception.
Kar said their approach helps identify the ANN components that truly match brain activity, supporting more reliable models for understanding how people see and interpret the world. He noted this is especially relevant to their autism research program, which uses models of the neurotypical brain as a baseline.
A neurotypical brain refers to patterns of development and functioning generally considered typical. Accurate baselines are important for studying how perception and information processing may differ in conditions such as autism.
To enable broader use, the team has released a publicly available testing toolkit that AI developers can apply to evaluate their models and find avenues for improvement.
Published in Nature Machine Intelligence, the study advances an evaluation approach that goes beyond task performance. The authors argue it is equally important to determine whether AI achieves results through processes that genuinely resemble those of the brain.
The work could guide AI development toward systems that not only recognize the world with high accuracy but also offer a truer picture of how the human brain interprets what it sees.