Vision models struggle to identify cultural context from museum images

Researchers introduced Appear2Meaning, a benchmark of 750 museum objects to test how vision-language models recover cultural metadata from images alone.

The benchmark evaluates how well artificial intelligence can identify a title, culture, period, origin, and creator based solely on visual input. Testing across nine vision-language models showed that mean field accuracy ranged from 36.16% to 49.82%.

Only a small fraction of records achieved full compatibility across all five categories, with some models reaching as low as 2.13% for a complete record match. The study also found that providing only an image resulted in higher reference compatibility for culture, period, and origin than a control group that received a museum description and a title. This suggests that visual data alone can sometimes provide more accurate cultural context than text-heavy museum records.