Multimodal AI Misreads Individual Historical Illustrations but Nails Group Patterns

A new EMNLP 2026 paper evaluated five multimodal AI models on nearly 2,000 historical children's book illustrations (1728–1932), comparing their interpretations with annotations from thousands of human volunteers. The models often erred on individual images yet achieved near-perfect accuracy at the group distribution level.

2026-11-23 ~ 2026-11-23 · 2 related posts