After people watched motion, their reported position of a stationary object shifted even though its image stayed unchanged. A study published Oct. 2 in Current Biology found a statistically significant leftward shift in object positions decoded from activity in macaque inferior temporal (IT) cortex after rightward motion exposure. After leftward motion exposure, decoded positions shifted right, but that change was not statistically significant. Tested AI vision networks did not reproduce motion-driven position shifts on their own.
A motion aftereffect is a perceptual change that follows sustained exposure to movement: after watching motion in one direction, a stationary target can seem to shift the other way. Because the target image stays unchanged, the illusion shows that reported position can shift even when the visual input does not.
In the human experiment, 22 participants viewed drifting gratings for 30 seconds, with three-second top-ups, before briefly seeing a stationary object image and reporting its center. After rightward adaptation, reports shifted left; after leftward adaptation, they shifted right, opposite to the preceding motion.
For macaque recordings, the researchers trained a position decoder on baseline activity in IT, then used it to estimate object locations from responses after adaptation. Decoded positions shifted left after rightward adaptation, a statistically significant result. After leftward adaptation, the estimate shifted right, but that change was not statistically significant. Together, decoded positions showed the same overall direction of change as human reports, although only the shift after rightward adaptation was statistically significant.
The team tested feedforward, recurrent and video-based neural networks. They represented object position, but their position representations did not shift after motion adaptation. In simulation, applying a transformation derived from macaque IT activity made the networks' position representations shift.
The experiment used a controlled laboratory illusion and a limited set of model families; it does not show that all AI vision systems fail at everyday perception. Technical details here are based on the authors’ March arXiv manuscript. Whether adding this kind of adaptation improves real-world performance remains to be tested.
Sources
- Yakubovskaya et al., “The macaque IT cortex but not current artificial vision networks encode object position in perceptually aligned coordinates,” Current Biology, DOI: 10.1016/j.cub.2026.09.019.
- Authors’ arXiv manuscript, version 1, submitted March 11, 2026.
- York University’s October 2 study announcement.
- Illustrative image: Cactus26 / Wikimedia Commons, CC BY-SA 3.0; the diagram is not from the study. Image details and license.
