A world model is an internal representation that lets an AI system simulate what will happen next — before it actually takes an action. Rather than blindly trying things and reacting to results, an agent with a world model can plan ahead by mentally "running" possible futures and picking the best path.

Humans do this constantly. Before merging on a highway, you don't just jerk the wheel — you mentally simulate the gap, the speed, the risk. Researchers want AI agents to have the same ability, especially for complex, multi-step tasks where mistakes are costly or hard to reverse.

The challenge is that world models can accumulate errors. Each simulated step is imperfect, and those small mistakes compound — a problem researchers call rollout error. This is an active area of AI research: how do you build agents that plan far into the future without their internal simulations drifting too far from reality?