Testing ChatGPT 5.6's Art Comprehension

kaljakin · reddit · 2026-07-13

This lengthy Reddit post tests ChatGPT 5.6's ability to judge abstract artistic emotions like "gentleness," "tranquility," and "energy" using a set of painting multiple-choice questions. The author found that the model not only looks at character actions and object relationships but also heavily utilizes visual cues like color, negative space, and brushstrokes.\n\nThe author originally intended to prove that the model "doesn't understand art," but after multiple tests, the model's answers and explanations were quite on point. It even quickly corrected itself when probed, supplementing with a seemingly self-consistent analysis. The post ultimately observes that multimodal models might possess a much stronger grasp of artistic semantics, emotion, and visual style than intuitively assumed.

Original post →

More from Multimodal

Multimodal channel →