World models: what the term covers

What a world model is

The term comes from reinforcement learning rather than from generative AI. In the agent-and-world picture formalised as a partially observable Markov decision process — the framing Sutton and Barto use throughout the standard textbook — an agent takes actions, those actions change the state of the world, and the agent never observes that state directly. What reaches it are observations. Something inside the agent therefore has to model how the state evolves, and the original use of “world model” belongs to that tradition.

Google DeepMind gives a working definition: world models are “AI systems that can use their understanding of the world to simulate aspects of it, enabling agents to predict both how an environment will evolve and how their actions will affect it”. The substrate differs from a language model. Where a language model learns the statistical structure of text, a world model learns the statistical structure of space and time — how light falls on a surface, how an object responds to force, what is present behind a wall that no camera has seen.

That difference decides what each kind of model is good for. A language model can describe a procedure; it cannot tell you what happens when the procedure is applied to a particular object in a particular place. A world model can predict the consequence and, by itself, cannot tell you whether the action was permitted.

One term, several different things

“World model” is one of the most overloaded terms in current AI. Computer vision, robotics, reinforcement learning and generative AI each claim to be building world models, and each means something different by it. A video model that renders beautiful but physically impossible flames, a language model improvising a playable game, and a physics engine that faithfully simulates combustion all go by the same name.

World Labs set out a functional taxonomy in June 2026 that separates the pieces by what they are for:

• Renderers produce what you see — the pixels of a scene, generated at speed and quality.
• Simulators predict how state changes under an action, whether it is physically plausible, and what follows from it.
• Planners predict in a compressed latent representation rather than in pixels, because a planner needs to know what state an action implies, not what a frame looks like.
• The loop connects them: observations in, actions out, state updated.

The practical consequence is that two systems both called world models may not be comparable at all. One may be a renderer optimised for visual fidelity, another a planner whose output is never meant to be looked at. The useful question is not which is better but which of the three functions a system implements, and on which substrate — pixels, or a learned representation.

What the dated record shows

The entries below are taken from each project’s own publication, which is also the limit of the claim: these describe what the teams state about their systems, not independently measured results.

Two bets are visible across the list. One predicts in pixels and is judged by how convincing the output looks. The other predicts in a learned representation and is judged by whether the predictions are useful for planning. The two camps disagree on compute cost, on what counts as understanding, and on which is the likelier route to general capability.

What a world model does not supply

A world model is a learned, non-symbolic representation of dynamics. Read that description for what is absent from it. There is no vocabulary, so nothing states what a thing is called. There is no provenance, so nothing states who asserted it or on what evidence. There is no permission model, so nothing states whether the agent may act on it. There is no currency, so nothing states whether the prediction still holds.

Those absences are the knowledge layer’s subject matter, and they are not incidental. A system that predicts the consequence of an action still needs something outside itself to decide whether the action was allowed, to record that it happened, and to check later whether the conditions still apply. World models extend what an agent can anticipate; they do not supply what it is entitled to do.

Fidelity is the second limit. Visual plausibility is not physical correctness — World Labs’ own essay gives the example of a video model producing gorgeous but physically impossible flames — and there is no settled measure of how well a world model models a world. Treat demonstration quality as a demonstration, and keep measured performance in a separate column.