← Back to all stories
TechnologyAI models unlike human brain

York University study finds AI vision models rely on internal processes unlike primate brains

Researchers at York University have shown that artificial neural networks widely used to model vision can predict primate brain activity but that brain responses poorly predict many of the models’ internal features. The asymmetry, absent when comparing brains to each other, indicates the AI systems solve object recognition partly via strategies the brain does not appear to use. The findings, published in Nature Machine Intelligence, introduce a reverse predictivity test and a public toolkit to improve model-brain alignment.

Key points

  • •AI models predict brain neuron responses well, but reverse prediction is weak.
  • •Brain-to-brain comparisons are symmetric, unlike model-to-brain ones.
  • •Aligned model parts better predict human behavior; toolkit released for testing.
21 November 20262 min read5 SourcesAI-generated — how does this work?

Why this is uncovered

Covered by ScienceDaily, university releases and specialist science outlets, with limited pickup in mainstream AI market or general news media.


⚠

This article was generated automatically from primary sources and has not been reviewed by a human editor. Verify claims before sharing.

A study from York University researchers has identified a key mismatch between how widely used artificial neural networks process visual information and how primate brains do so. Artificial neural networks (ANNs), computer systems trained for tasks such as object recognition in images, have become leading tools in neuroscience for explaining the brain’s visual system because they can reproduce patterns of neural activity.

However, according to the research led by Assistant Professor Kohitij Kar, Canada Research Chair in Visual Neuroscience, and postdoctoral fellow Sabine Muzellec, success in one direction of prediction does not confirm matching internal mechanisms. “Artificial intelligence systems are often described as ‘brain-like’ because they can predict activity in parts of the brain that help us recognize objects,” Kar said in a York University release. “Until now, scientists mostly tested this in one direction. They asked whether AI models can predict brain activity.”

The team developed a reverse predictivity test. If the models truly mirrored the brain, recorded neural activity should also predict the models’ internal features with comparable accuracy. They examined responses to 1,320 natural and synthetic images of familiar objects—a bear, elephant, face, apple, car, dog, chair, plane, bird and zebra—set against varied backgrounds, plus 300 additional images of the same objects in outlines, drawings and other styles (York University).

Results showed a clear asymmetry. AI models predicted the responses of recorded neurons in macaque inferior temporal (IT) cortex reasonably well, explaining roughly 50 percent of variance in many cases. Yet the same neural responses were far less effective at predicting many of the models’ internal units. “The results were striking. While AI models can predict the neurons we recorded in the brain fairly well, the brain cannot equally predict many of the model’s internal features. Interestingly, this is not the case when neurons from one brain is compared against ones from another brain,” Kar stated (ScienceDaily).

Monkey-to-monkey mappings remained symmetric, indicating the mismatch is specific to the models rather than a general limitation of neural data. The findings suggest today’s AI systems solve visual tasks partly using internal strategies the primate brain may not employ. Parts of the models that do align with brain activity also proved better at predicting real human behavior (Medical Xpress).

The study, published in Nature Machine Intelligence, notes that continued refinement of models solely on forward predictivity could widen differences over time, reducing their value as hypotheses for biological vision. Similar ANNs are used to study auditory, language and motor systems, raising parallel concerns. Muzellec said the work “challenge[s] how similar current AI systems really are with the primate brain” and supplies “a well vetted diagnostic metric for the field” (York University).

The researchers released a public testing toolkit so developers can evaluate reverse predictivity and identify common units shared with IT cortex versus unique units lacking biological counterparts. Improving alignment could strengthen models used in behavioral experiments and clinical research, including autism studies that rely on neurotypical baselines. The paper is available at Nature (Nature Machine Intelligence).

Share:X / TwitterLinkedIn

Want to respond?

The Uncovered has no comment section by design — we don't host or moderate discussions. If you still want to read along or join in, you can do so over Nostr, an open protocol where your replies live on your own key instead of on our server. Install NostrComments for Chrome or Firefox and the discussion appears on top of this page. Works on desktop and on Firefox for Android.

Open source, no tracking, free.

More in Technology

IEA projects global data centre electricity use to double by 2030 on AI growth

The International Energy Agency forecasts that electricity consumption by data centres worldwide will roughly double by 2030, reaching around 945-1,000 TWh, driven primarily by artificial intelligence. The United States and China are expected to account for the bulk of the increase, with implications for power grids and generation mixes. Related carbon emissions from the sector are projected to rise significantly in the near term.

20 Nov 2026

Scientists develop seismic method to detect buried water ice on the Moon

Researchers from the University of Maryland, Lawrence Berkeley National Laboratory and the University of Hawaii have shown that seismic waves from moonquakes can locate and map subsurface water ice deposits. Lab experiments, temperature modeling and simulations indicate ice stiffens lunar soil, speeding waves two to three times and creating detectable reflections. The approach could aid Artemis and other missions seeking resources at the lunar south pole.

7 Aug 2026