When a robot moves an object or a self-driving vehicle approaches an intersection, it needs to predict what will happen after it acts. In artificial intelligence, a world model is an internal representation of an environment that helps an agent predict possible outcomes and compare actions before acting.
World models matter because they give AI a way to plan around changing physical conditions. Robots, vehicles, and other embodied systems can use them to test decisions in simulation. Their predictions are only as reliable as the model, however: errors in the imagined world can lead to poor decisions in the real one.
How a world model supports planning
A world model learns patterns in how an environment changes. It might represent objects, movement, and cause and effect, then predict what the environment will look like after an action. An agent can compare possible moves with those predictions and choose one that helps it reach its goal.
For example, a robot tasked with moving a cup could predict whether pushing it will make it slide, tip, or collide with another object. It can then choose a movement that is less likely to spill the cup. This trial and error in the model can reduce the need to test each action in the physical environment.
Some world models compress what they observe into a smaller representation, rather than trying to reproduce every detail. In the original World Models research, researchers trained agents in simulated environments using learned representations of those environments. The approach shows how an agent can practice inside a model’s “dream” and use those simulated experiences to guide its behavior.
How world models differ from language models
Large language models learn patterns in text and generate language in response to prompts. World models focus on predicting how an environment changes, especially in response to actions. The two approaches can complement each other: a language model can interpret an instruction such as “move the cup to the table,” while a world model predicts the physical consequences of possible movements.
Researchers disagree about how much physical or causal understanding language models acquire from text. Some argue that they form useful internal representations of the world; others argue that those representations remain incomplete or rely heavily on learned patterns. The question remains contested, as discussed in a survey of views on language models and world models.
One research direction, Yann LeCun’s Joint Embedding Predictive Architecture, or JEPA, predicts relationships between learned representations rather than recreating every detail of an input. The architecture uses a context encoder, a target encoder, and a predictor. This approach aims to learn useful structure from observations without generating a pixel-perfect copy of the scene, as described in research on world models and AI.
Where world models are being used
World models help AI agents predict consequences when they act. Robotics is a direct example: a robot can use predictions to plan how to grasp, move, or avoid objects. The same idea supports research in autonomous driving, model-based reinforcement learning, and interactive simulation.
A different use is generating environments that people or agents can explore. Google DeepMind describes Genie 3 as a general-purpose world model that creates interactive environments from text prompts. Its environments support real-time navigation at 24 frames per second. This system shows how world models can support interactive simulation. A convincing generated environment alone does not show that an AI can reliably reason about the physical world.
What world models still get wrong
A world model can make planning more efficient, but its predictions can drift from reality. If small errors build up across a sequence of predicted actions, the agent may choose a plan that works in simulation and fails in the real environment. An agent can also exploit flaws in its model, finding a simulated action that appears to meet its goal but produces an unintended result outside the simulation. Research and implementation guidance also describe training instability and generated outcomes that become less varied, a problem known as mode collapse.
World models also raise data questions. Robots and other physical AI systems learn from sensor data about real places and objects. Collecting and using that data can raise privacy and copyright concerns, while synthetic data does not automatically settle who owns the underlying material or whether a digital reconstruction reproduces protected content. These questions are discussed in the context of physical AI and legal risks.
To evaluate a world model, check whether it predicts the outcomes that matter for the task. A model used to move fragile objects needs to predict contact and breakage well enough to support safe decisions; visual realism alone is not a sufficient test. The broader claim that world models will deliver more efficient or general intelligence remains unproven.