Summary:
A new study published in Current Biology by researchers at York University reveals a core computational difference between biological vision and contemporary artificial intelligence. By testing the classic motion aftereffect illusion in human observers, macaque monkeys, and several leading artificial neural networks, the team found that primate inferior temporal (IT) cortex represents object position in a way that shifts with recent visual history, matching subjective perception. In contrast, current AI vision models maintain rigid, pixel-based representations that do not reflect these history-dependent perceptual shifts. These results introduce a practical benchmark for NeuroAI and suggest that human-compatible machine vision must incorporate dynamic, history-dependent computations to align with biological perception.
Key Facts:
- The Illusion Divide: After adapting to directional motion, human observers perceive a stationary object as displaced in the opposite direction. Neuronal population codes in primate IT cortex shift in the same perceptual direction, whereas state-of-the-art AI vision networks do not reproduce this effect.
- History-Dependent Perception: Biological vision continually updates spatial representations based on recent temporal experience, treating perceptual shifts as adaptive computations. Artificial systems tested remain strictly tied to objective pixel coordinates.
- New NeuroAI Benchmark: The study highlights that training AI exclusively for static pixel accuracy misses essential dynamical computations. Incorporating adaptation-driven, history-sensitive mechanisms may be necessary for safe, intuitive human–machine alignment.
Source: York University
Biological vision does more than record light intensity; it actively integrates recent visual history to construct a usable, adaptive perception of the world.
Rather than functioning like passive digital cameras that continuously log raw pixel coordinates, human and primate visual systems use adaptive computations to interpret dynamic scenes. These computations can produce systematic perceptual biases. In the motion aftereffect illusion, prolonged exposure to motion in one direction causes a subsequently stationary object to appear shifted in the opposite direction. Although the retinal image remains unchanged, conscious perception of position is altered.
Neuroscientists increasingly interpret such perceptual “errors” as signatures of efficient, adaptive processing rather than flaws. The York University team asked whether modern artificial vision models, intended to interact with human users and interpret visual scenes, reproduce the same history-dependent perceptual computations.
Published in Current Biology, the study shows that leading artificial vision networks lack the temporal flexibility observed in biological systems. While many deep neural networks accurately classify objects in static images, they evaluate inputs primarily through feedforward, pixel-based computations and do not dynamically reshape spatial codes based on recent visual context.
The research, led by graduate student and first author Elizaveta Yakubovskaya, combined human psychophysics with intracortical electrophysiological recordings from macaque IT cortex, a high-level visual area implicated in object recognition. Human participants and macaque subjects were exposed to motion adaptation and then presented with stationary targets designed to elicit the position-shift illusion.
Human observers reliably reported the illusion. Electrophysiological data revealed that IT population representations of object position shifted in the same direction as the perceived displacement, despite unchanged retinal input. When the same experimental paradigm was applied to several standard artificial vision networks—including feedforward, recurrent, and video-based models—none exhibited the adaptation-induced position shifts observed in primates. The networks continued to represent object position according to objective pixel coordinates.
Bridging the NeuroAI Gap for Human–Machine Alignment
These findings suggest that primate IT does more than encode object identity: it encodes position in a perceptually aligned coordinate frame that reflects recent sensory history. This reveals a missing computational ingredient in current artificial vision systems—adaptation-driven representational warping—that links neural dynamics to subjective perception.
The gap has practical implications. As autonomous systems and assistive technologies operate alongside humans—whether in driving environments, clinical imaging, or interactive robotics—mismatches in how space and motion are perceived could produce subtle but consequential misalignments. Building AI that reliably collaborates with people requires not only correct output labels but also a deeper match to the computations that produce human perception and behavior.
“Today’s AI vision systems are impressive, but they still do not always see the world the way we do,” said senior author Kohitij Kar, Ph.D., Canada Research Chair in Visual Neuroscience and Assistant Professor at York University. “By combining neuroscience experiments with AI, we can identify computations biological vision uses and incorporate them into machine systems to make them more brain-like and human-compatible.”
Editorial Notes:
- Article content edited for clarity and context.
- Original journal paper was reviewed in full by the authors of this summary.
- Additional explanatory material was added by the editorial team to aid reader understanding.
About this AI and visual neuroscience research:
- Media Contact: Sandra McLean
- Source: York University
- Image Credit: Image credited to Neuroscience News
- Original Research (Open Access): Current Biology (Sept 22, 2026). Title: “The macaque IT cortex but not current artificial vision networks encode object position in perceptually aligned coordinates.” Authors: Elizaveta Yakubovskaya, Hamidreza Ramezanpour, Matteo Dunnhofer, and Kohitij Kar.
- DOI: 10.1016/j.cub.2026.09.019
Abstract
The macaque IT cortex but not current artificial vision networks encode object position in perceptually aligned coordinates
Efficient interaction with the visual world requires both object identification and accurate localization. Although spatial “where” processing has traditionally been associated with dorsal stream pathways, recent work shows that ventral stream areas, including IT cortex, also carry position information. Because many experiments couple position to pixel-based retinal coordinates, it has been unclear whether ventral stream position signals reflect perceptually meaningful coding or simply inherit retinotopic inputs.
This study uses the motion aftereffect to dissociate perceived position from retinal location while keeping visual input constant. Combining intracortical recordings in macaque IT with matched human psychophysics, the authors demonstrate that motion adaptation produces direction-opponent biases in IT population codes that mirror human perceptual reports. Motion adaptation reshapes IT representational geometry in a way that likely contributes to perceptual biases.
Extending the analysis to artificial systems, the researchers show that feedforward, recurrent, and state-of-the-art video models fail to exhibit adaptation-induced position shifts, despite accurately encoding pixel-based object position. Importantly, applying empirically derived IT-style transformations to model features can reproduce the perceptual effect, indicating that adaptation-driven representational warping is a missing computational mechanism in current artificial vision.
Together, these results identify IT as a candidate locus for perceptually aligned spatial coding, propose adaptation-driven restructuring as a mechanism linking neural dynamics to perception, and highlight a principled gap between biological and artificial vision that NeuroAI should address.
