AI Still Isn’t as Good at Recognizing Objects as People Are, New Test Shows
A study by Mugihiko Kato and Biyu J. He of New York University, published in iScience on 17 September 2026, built an image set that systematically untangles global shape, internal parts and texture, then compared human viewers against more than 200 deep neural networks. No model reproduced the human cue-reliance profile, and every model tested substantially underperformed people when global shape was the only usable cue. This article covers the design, the finding that brain-alignment scores did not predict behavioural alignment, which model families came closest, and what the result means for autonomous driving, robotics and assistive devices.