AI Vision Systems Fail Basic Illusion Test That Primate Brains Pass, York University Study Reveals

AI Vision Stumbles Where Primate Brains Succeed

A York University study published in Current Biology exposes a critical blind spot in AI vision systems: they cannot reproduce basic perceptual illusions that human and primate brains experience naturally1

1

. The research team, led by Kohitij Kar, Canada Research Chair in Visual Neuroscience, tested how biological vision and artificial neural networks respond to the motion aftereffect illusion. When humans stare at continuous directional motion and then view a stationary object, they perceive it as displaced in the opposite direction—even though the physical light hitting the retina hasn't shifted2

2

. This gap between biological and AI vision reveals fundamental differences in how each system processes spatial information.

History-Dependent Vision Versus Pixel-Perfect Accuracy

Biological vision doesn't operate like a passive digital camera. First author and graduate researcher Elizaveta Yakubovskaya explains that when human observers and macaque monkeys were exposed to motion adaptation, both reported the position-shift illusion, and neural representations in the primate inferior temporal cortex shifted accordingly1

1

. The inferior temporal cortex dynamically reshapes spatial representations based on recent temporal experience, treating these perceptual shifts as adaptive features rather than flaws. Current state-of-the-art artificial vision networks, however, remained strictly tethered to static pixel coordinates, showing no adaptation whatsoever2

2

. While deep artificial neural networks excel at spatial object recognition in static frames, they evaluate each image primarily through feedforward processing without the continuous temporal adaptation characteristic of animal vision.

Building Human-Compatible Machine Vision

The research establishes a new NeuroAI benchmark for evaluating dynamic vision models. "Today's AI vision systems are impressive, but they still do not always see the world the way we do," says Kohitij Kar2

2

. The findings demonstrate that building AI systems that work safely and intuitively alongside humans requires more than getting the right answer—it demands training them on human-like perceptual computations rather than solely optimizing for static pixel accuracy. The York team combined human psychophysics with electrophysiological recordings from primate brains to pinpoint exactly where artificial and biological vision diverge1

1

.

What This Means for AI Development

Neuroscientists increasingly view perceptual "mistakes" like the motion aftereffect illusion as signatures of optimal, energy-efficient biological computations. Kar emphasizes that if we want AI that works with humans and understands the world in more human-compatible ways, we cannot focus only on whether it gets the right answer2

2

. The study captures the promise of NeuroAI—bringing neuroscience and artificial intelligence together to reveal the computations biological vision uses that AI still lacks. By leveraging motion adaptation to show where perceived and pixel-based positions diverge, the researchers created a benchmark to evaluate whether increasingly capable AI vision systems will become more like us or increasingly different from us. The findings not only advance understanding of the inferior temporal cortex's role in spatial information encoding but also provide actionable insights for building better, more brain-like artificial systems that incorporate visual history into their processing.