AI describes art differently to different audiences — but barely improves on its own

Researchers built a test to see whether language models can describe artworks in ways that actually help people from different cultural backgrounds understand them. It turns out the base models barely manage the task, and even with tweaks that consider how listeners absorb information, improvements are small — around 8 percent better comprehension in human testing.

This is a measurement problem dressed as a cultural one. For years, AI developers have assumed that if a model knows facts about art and cultures, it will naturally explain them well to different audiences. This paper shows the assumption is wrong — knowing what something means and explaining it so a specific person understands it are different tasks. The gap matters because cultural bias in AI systems usually gets framed as a missing dataset problem. This suggests it's actually a comprehension problem: the model doesn't know how to translate between what it knows and what a particular person needs to hear.