AIs Should Learn Human Vision Quirks

A York University study found humans (and macaques) exhibit a tiny but consistent motion aftereffect that state‑of‑the‑art AI vision networks do not reproduce. The authors argue that if we want AI systems that interact seamlessly with people (in AR, driver assistance, medical displays), training them to reproduce human perceptual errors — not only accuracy — could improve compatibility and neuroscientific modeling. — This reframes AI design tradeoffs: accuracy alone may be the wrong optimization when human compatibility, safety, and manipulation risk depend on shared perceptual computations.