A world model is an AI model that learns an internal representation of how an environment behaves so that, given the current state and an action, it can predict or simulate what comes next. The goal is to capture the rules of an environment—object motion, physics, spatial structure—from data. The concept comes from reinforcement learning, where an agent plans and practices inside a learned model instead of the real environment, and was popularized by the 2018 paper "World Models" by David Ha and Jürgen Schmidhuber.
More recently the term also refers to models trained on video that generate 3D environments in real time in response to user input, with Google DeepMind's Genie series as a prominent example. World models are used for simulation in robotics and autonomous driving, for generating game environments, and for creating training environments for AI agents. In contrast to language models that predict the next token of text, they are often discussed as a path toward AI that understands the physical world.