AI vision models can't understand physics — researchers add simulation to fix it

Current AI models that process images and video fail at basic physics tasks like predicting what happens next in a fluid or watching whether objects move realistically. A research team built a method that feeds the models physics simulation data during training, and it cut their error rate by up to 20 percent on fluid dynamics tasks.

This is a clean demonstration of what AI vision models are actually missing: they can describe what they see in an image, but they don't understand how the physical world behaves. The fix is straightforward — embed real physics into training — which means that multimodal AI systems designed to reason about the real world are currently reasoning about images, not physics. The gap matters if you're using these models for anything involving prediction or planning in physical space.