Scientists used to think artificial neural networks (ANNs) mimicked the primate visual cortex fairly well.
If an AI model can predict which neurons fire when you see a cat, researchers assumed it was processing that image in roughly the same way you are. It seemed like a reasonable assumption.
York University researchers didn’t buy it.
They turned the table. Instead of asking if AI predicts the brain, they asked if the brain predicts AI.
This simple shift—reverse predictivity in neural models —reveals a messy truth. The internal wiring of top-tier AI vision models doesn’t just differ slightly from biological brains. It relies on strategies the brain simply does not use.
Testing the mirror: why the standard AI validation is flawed
For a decade, the go-to metric for how brain-like is an AI vision model has been unidirectional. Scientists feed images to both a human participant (measured via brain scans or recordings) and a computer. If the computer’s output aligns with the neural spikes, the model is “good.”
Kohitij Kar, an assistant professor at York University, notes that this only tests half the relationship.
“Artificial intelligence systems are often ‘brain-like’ because they predict brain activity,” Kar explains. “Until now, we tested it one way.”
To fix this blind spot, Kar and his team, including postdoctoral fellow Sabine Muzellec, designed a bidirectional test. If the connection is symmetric—as a true mirror should be—recording brain activity should be able to predict the internal activations of the AI model just as well as the AI predicts the brain.
They ran this reverse prediction test on a massive dataset:
- 1,320 realistic images of objects (bears, chairs, planes, zebras).
- 300 stylistic variations, including outlines, sketches, and abstracted forms.
The goal wasn’t just to see if the AI could “recognize” the objects. It was to see if its thought process —its hidden layer activations—mapped to the primate’s neural processing.
The asymmetry: why AI answers aren’t like brain answers
The results broke the mirror.
When AI models predict brain activity? Decent match.
When brain activity predicts AI’s internal features? Not even close.
“This is striking,” says Kar. “While AI models can predict neurons fairly well, the brain cannot predict the model’s features. This mismatch doesn’t happen when we compare brains to brains.”
What does this mean for which AI models actually simulate human vision?
It suggests ANNs reach the right answer (“That’s a dog”) through a completely different computational route than a primate does. They are getting to the same destination, but driving a very different car.
This matters because many fields—neuroscience, psychiatry, even clinical diagnostics for conditions like PTSD or autism—rely on these models as baselines. If you’re trying to understand neurotypical processing to identify deviations in autism spectrum disorder, you need a model that actually thinks like the baseline subject. Not one that cheats via visual heuristics the brain ignores.
Why “brainlike” claims are risky for research
Using poorly aligned AI systems carries real risks.
Much of modern behavioral research assumes AI processes visual information similarly to humans. This study shows that assumption is wrong.
Muzellec warns that these “brainlike” models rely on internal components the brain likely abandons long ago in evolutionary development.
“Our findings challenge how similar current systems really are… We show that models thought to be brainlike rely on internal components the brain does not appear to leave trace of.”
So, which parts of the AI are actually useful?
Surprisingly, the few components of the ANNs that did align with neural activity were the same ones best at predicting human behavior.
This gives researchers a new filter. You can now sort AI layers:
1. Those that mirror the brain (and likely help explain human perception).
2. Those that rely on alien, non-biological strategies (which should be discarded when studying human cognition).
How this changes future neuroscience
This isn’t just theoretical. The team published a diagnostic metric in Nature Machine Intelligence in March 2025 (note: original text cited 2026/01204-0 DOI format suggests near-future/current timeline).
They’ve released a toolkit for developers.
Now, you don’t just ask if the model recognizes the bear. You ask if a bear-recognizing neuron pattern can reverse-engineer the model’s thought process.
If not? You’ve got a model that passes the Turing Test but fails the Biology Test.
As ANNs expand into language, hearing, and movement processing, this validation method becomes essential. We might be building millions of parameters on foundations that don’t shake hands with biology.
The toolkit is out there.
But it raises an uncomfortable question. If the best AI models are structurally so different from our own brains, does that mean we’ve been modeling our minds all wrong?
Or does it just mean we’re smarter than we thought—and our computers are clever liars?
