Multimodal AI models like Gemini 3 Pro perform poorly on a new basic vision test (WorldVQA), failing to crack 50% accuracy.

Source Brief

Multimodal AI models like Gemini 3 Pro perform poorly on a new basic vision test (WorldVQA), failing to crack 50% accuracy. The benchmark tests whether models actually recognize objects in images or just guess. This highlights a significant gap between impressive demo capabilities and AI's actual understanding of the world.