AI Video Models Attempt to Mimic Real Physics Without Understanding It

AI video models attempt to mimic real-world physics, yet they lack an actual understanding of it.

The Limits of Artificial Intelligence in Understanding Physical Laws

Recent findings reveal that AI video generators fail to grasp fundamental physical laws by merely observing videos. This challenges the effectiveness of these models to autonomously learn from visual data without human intervention.

A Revealing Study on AI Capabilities

Researchers from Bytedance Research, Tsinghua University, and Technion tested these models by creating a 2D simulation with simple shapes and motions, generating hundreds of thousands of mini-videos for training and testing. The results showed that the models could “imitate” physical laws, but not understand them.

Moreover, when faced with new and unexpected situations, the models failed to respond accurately. At best, they attempted to replicate the closest training example available.

A Clear Hierarchy in Model Priorities

The researchers also noticed that the video generator often transformed one shape into another (e.g., a square randomly turning into a ball) or made other illogical adjustments. It became clear that the model’s priorities follow a distinct hierarchy, with color being the most significant, followed by size, then speed, with shape being the least important.

A Challenge for the AI Community

“It’s hard to determine whether a video model has learned a law or just memorized data,” stated Bingyi Kang, the lead author of the study. This highlights that AI relies more on training examples than on learning universal rules, regardless of data volume. Addressing this issue remains a challenge for the entire AI community.