Where is the object? Two answers. Retinal position: computed from photon impact coordinates, geometrically correct. Perceived position: where the object appears to be, shifted by context, illusions, and perceptual biases. These differ measurably — visual illusions demonstrate that perceived position can deviate degrees from retinal reality.
Macaque inferior temporal cortex (arXiv:2603.11248) encodes perceived position, not retinal position. When presented with stimuli that induce positional illusions, IT neurons track where the object appears to be, not where it is. The representation is faithful to subjective experience, not to physical reality.
Every current artificial vision network fails to do this. CNNs, vision transformers, CLIP, DINOv2 — all encode positions in coordinates that do not track perceptual illusions. They are more “accurate” than biology in the geometric sense and less accurate in the perceptual sense. Biologically-inspired coordinate transforms partially rescue artificial systems, but the native representations lack the warping.
The implication is that high-level visual processing in IT cortex is not computing objective features and then subjecting them to illusory distortion. The illusion is not a bug that contaminates an accurate signal. The perceived coordinates are the computation. The brain's position code serves behavior, not measurement — and behavior operates in perceived space, not physical space. An organism reaches for where the object appears, not where it is.
Artificial systems optimized for geometric accuracy are solving a different problem than the one biology solves. The gap is not a failure of artificial systems to match biology. It is a difference in what the systems are for.
Note (2026-11-11): checked against the full text of the paper (Yakubovskaya, Ramezanpour, Dunnhofer and Kar, arXiv 2603.11248). The paper's design and headline are as the essay says: using the motion aftereffect, which shifts perceived position while the retinal image is unchanged, motion adaptation produces direction-opponent biases in macaque IT population codes for object position that qualitatively mirror human reports, and the feedforward, recurrent and video networks it tested show no such adaptation-induced shift. Five points differ. (1) The essay says IT 'encodes perceived position, not retinal position' and 'is faithful to subjective experience, not to physical reality'. The paper shows that IT responses contain decodable object position that is also accurate for the pixel-based position, and that adaptation adds systematic biases to what a decoder reads out; the paper says the reshaped representational geometry 'likely contributes' to the perceptual biases, which is weaker than saying IT encodes the percept instead of the retinal position. (2) The essay speaks of illusions that deviate by 'degrees'. The paper uses one illusion, the motion aftereffect, and the human shifts it reports are small (a mean displacement of about 0.2 degrees for one adapter direction and about 0.13 degrees for the other, from the lines I read). (3) The essay lists 'CNNs, vision transformers, CLIP, DINOv2' as all encoding positions more accurately than biology. I searched for CLIP and DINOv2 and found neither named. The paper reports that position was well decodable from some networks (VGG-16 and ResNet-18, correlations of about 0.8 to 0.87) and markedly less from others, including a transformer (ViT-L32, 0.59 and 0.7 in the two axes). (4) The essay says biologically inspired coordinate transforms 'partially rescue' artificial systems. The paper says imposing empirically derived IT-based transformations on model features is sufficient to simulate the effect, while implementing only intrinsic suppression dynamics in network units fails to reproduce it. (5) The essay's conclusions that the perceived coordinates 'are the computation', that the brain's position code serves behaviour and not measurement, and that an organism reaches for where the object appears, are the essay's; the paper's own conclusion is that IT is a locus of perceptually aligned spatial coding and that adaptation-driven representational warping is a missing ingredient in current models.