Overview
- A NYU Grossman School of Medicine study published Sept. 20, 2026 compared human observers with more than 200 deep neural networks and found a consistent gap when only an object’s silhouette was available.
- The team tested 240 manipulated images from 48 everyday categories to separate global form, internal parts and surface texture so they could pinpoint which visual cues each observer used.
- In a striking experiment the researchers filled outlines with many small cross marks and human participants still named objects by contour while most AI models misclassified them as patterns like crossword grids or window screens.
- Models fine‑tuned with contrastive image‑text training matched human choices more closely on ordinary images but lost that advantage when global shape cues were disrupted.
- The authors propose a new diagnostic evaluation and urge training changes to build shape‑sensitive vision systems because robotics, autonomous driving and visual prostheses need reliable silhouette recognition under low visibility.