Warum KI-Videos die Physik falsch verstehen

Video generators are trained on huge collections of footage to predict or denoise frames. The training objective rewards output that resembles real video. It does not reward output that obeys conservation of mass or momentum. Common physics, such as a ball falling or a person walking, is well represented in the data and usually looks right. Rare or complex events, such as glass shattering, a specific collision, or liquid pouring into an oddly shaped container, appear less often and are harder to learn. OpenAI's own technical material on Sora noted that it did not accurately model interactions like glass shattering. Object permanence is a memory problem. When something is hidden, the model has to carry its identity through frames where it is invisible. Models attend over a limited amount of context, so hidden objects can come back changed or not at all. Cause and effect is also weak: a bite, a spill or a dent should leave a lasting change, but the model may treat it as a passing visual event. Clip length is limited mainly by compute. A video is represented as spacetime tokens, and attention cost rises steeply as the token count grows, so many products generate seconds rather than minutes. Longer videos are often made by extending clips, which lets errors compound. A common misconception is that these models contain a physics engine. They do not, though they may pick up some implicit physical regularities. Researchers have built benchmarks such as Physics-IQ to test this and found that visually realistic output does not mean physical understanding. World-model research tries to close the gap: Google DeepMind's Genie line generates interactive environments, Meta's V-JEPA learns by predicting in an abstract representation space instead of pixels, and NVIDIA's Cosmos targets physical AI such as robotics.