1. Home
  2. Science and Technology
  3. Does AI function like the human brain? New test shows machines use different shortcuts to recognize objects
Leave a comment 3 min of reading

Does AI function like the human brain? New test shows machines use different shortcuts to recognize objects

Author profile image Romário Pereira de Carvalho
Written by Romário Pereira de Carvalho Published on 22/11/2026 at 14:03 Updated on 22/11/2026 at 14:04
Google News
Be the first to react!
React to this article
Prefer CPG on Google

A study published in Nature Machine Intelligence evaluated 1,620 images and found that artificial neural networks imitate brain responses only on the surface, using internal strategies absent from human biology to identify everyday objects

Researchers at the University of York applied a first-of-its-kind test to 1,620 images and uncovered a flaw in comparisons between AI and the human brain. Although artificial systems predict how neurons respond when recognizing objects, actual brain activity cannot predict how these machines work internally.

How the reverse prediction test worked

For a decade, neuroscientists have used artificial neural networks, programs designed to identify patterns and objects in photos, to study biological vision. The premise was simple: if a program imitates neural responses, the two work in the same way.

A team led by professor Kohitij Kar and postdoctoral researcher Sabine Muzellec decided to reverse that logic. They developed what they called a reverse prediction test to check whether the relationship between machines and biology held in both directions.

Instead of simply observing whether a computer could predict the behavior of brain cells, the scientists did the opposite.

They measured whether real signals recorded from neurons could explain what was happening inside computer systems.

The experiment evaluated 1,320 natural photographs and realistic synthetic images featuring ten familiar subjects: a bear, elephant, face, apple, car, dog, chair, airplane, bird, and zebra. The items appeared in a range of indoor and outdoor settings.

The authors also added 300 more images featuring the same ten objects in different forms, including outlines, drawings, simplified representations, and artistic versions. The study was published in the scientific journal Nature Machine Intelligence.

The flaw found in vision models

The results showed that the similarity between machines and biology exists only on the surface. The programs successfully predicted neuron activity. Neurons are cells that transmit electrical and chemical signals through the nervous system.

In the reverse direction, however, biological data failed to predict many of the machines’ internal mechanisms. When the scientists compared neurons from two different brains, this failure did not appear, demonstrating that the problem lies with the algorithms.

“Although AI models can predict the neurons we record in the brain quite accurately, the brain cannot predict many of the model’s internal features with the same degree of accuracy“, Kohitij Kar explains.

In practice, the computer reaches the same visual result using shortcuts and internal strategies that biology does not use. The study also found that only parts of AI systems aligned with neurons accurately predict human behavior.

Direct impact on medical and clinical research

The discovery calls for caution among scientists who use computer programs to plan experiments and investigate neurological or psychological disorders. Many studies start from the assumption that computers process data exactly as people do.

If the underlying technology works differently, conclusions about human behavior could be distorted. Models once viewed as faithful copies, Sabine Muzellec says, rely on internal components absent from primate brains.

The warning also extends to other areas of neuroscience. Similar algorithms already support research into hearing, language, and body movement, fields that now need to reassess how much these tools really explain about the nervous system.

Open tool and next steps in science

The researchers are not calling for artificial neural networks to be abandoned, but for them to be fine-tuned. By identifying which parts of the code align with biological signals, researchers can build programs that more faithfully reflect reality.

Kar’s group applies this improvement to autism research. The work uses models of the neurotypical brain, a term describing neurological patterns considered common in the population, as a reference for understanding differences in perception.

With more accurate systems, scientists hope to advance research into conditions such as autism and post-traumatic stress disorder. The focus shifts from simply getting the answer right to also examining the method used.

To speed this shift, the University of York team has made a public suite of testing tools freely available. Developers around the world can now download the materials to audit their own vision systems.

Source: Yorku.

Sign up
Notify of
guest
0 Comments
most recent
older Most voted
Romário Pereira de Carvalho

I have published thousands of articles on recognized portals, always focusing on informative, direct content that provides value to the reader. Feel free to send suggestions or questions.

Share in apps
Download app
0
I'd love to hear your opinion, please comment.x