A model of what happens next
Most AI systems that act in an environment are built around a policy: a function that takes in the current state and directly outputs an action. A world model takes a different approach. Instead of mapping straight from state to action, it learns to predict how the environment itself will change, given the current state and a proposed action. In other words, it learns a model of "what happens next," separate from any decision about what to do about it.
This distinction matters because a world model is not, by itself, a way of choosing actions. It's a simulator the agent carries around in its own head. An agent that has learned an accurate world model can ask hypothetical questions before committing to anything in the real world: what would happen if I pushed this object, turned this valve, or took this step. A policy alone can't answer that question, it can only tell you what it would do.