AI reads 2,000 historical illustrations: individual errors, near-perfect aggregate patterns

_akpiper · x · 2026-11-23

An EMNLP 2026 paper tests GPT-4.1, Gemini 2.5 Flash, Claude Sonnet 4.5, Qwen3-VL-8B and LLaVA-13B on 2,000 historical children's book illustrations against 700k+ crowd annotations. Key findings: 97–98% accuracy on illustration detection; models systematically underestimate emotional intensity; and crucially, even with individual errors, models reproduce corpus-level feature distributions with correlations up to 0.99 — making them viable tools for large-scale visual culture research.

Related event: Multimodal AI Misreads Individual Historical Illustrations but Nails Group Patterns(2 posts)→

Original post →

More from Multimodal

Multimodal channel →