World model
In short
A world model is an AI system that learns how an environment behaves, so it can predict what happens next, including how the scene changes when someone acts in it. It lets AI simulate a world instead of only describing it.
- Also called
- World models, Generative world model
- Topic
- How AI models work
- Level
- Technical
- First described
- David Ha and Jürgen Schmidhuber, "World Models", March 2018. Simulating an environment to plan is an older idea, but this paper brought the term into mainstream machine learning1.
A chatbot predicts the next word. A world model predicts the next state of an environment: the next video frame of a road, or what a game level looks like after the player turns left. Google DeepMind describes world models as systems that use their understanding of the world to simulate aspects of it, so an agent can predict how an environment will evolve and how its actions will affect it2.
The idea was made widely known in machine learning by David Ha and Jürgen Schmidhuber's 2018 paper "World Models". They trained a generative network to learn a compressed picture of a video game environment over space and time, then trained a small agent using that model's features. The agent could even be trained entirely inside the model's own simulated "dream" and still work in the real game1.
Recent world models go much further. DeepMind's Genie 3, announced in August 2025, turns a text prompt into an interactive environment a person can move through in real time, at 720p and 24 frames per second, staying consistent for a few minutes2. It was released as a limited research preview, not a public product2.
A world model is not the same as a video generator, though they overlap. The key difference is control: a world model responds to actions, such as steering or walking, rather than playing out a fixed clip.
Self-driving is one active use. The team behind GAIA-1 says a model that generates realistic driving scenarios can speed up the training of autonomous driving software3.
Example
Wayve, a London-based self-driving company4, built GAIA-1, a world model that takes video, text and driving actions and generates realistic driving scenes, so its software can be trained on situations that are rare on real roads3.
Why it matters
World models move AI from describing the world to simulating it, which is what planning and acting need. They also let cars and robots practise in simulation before going near people.
Questions people ask
What is a world model in AI?
An AI system that learns how an environment works so it can predict what happens next, including the effect of actions taken in it. Google DeepMind describes world models as systems that simulate aspects of the world2.
Can I use Genie 3?
Only in limited form. DeepMind launched Genie 3 as a research preview for a small group of academics and creators, and said it was exploring wider access later2.
Is a world model the same as a video generator?
Not quite. Both produce moving images, but a world model reacts to actions as it goes. Wayve's GAIA-1, for example, takes driving actions as input and gives fine control over the vehicle's behaviour in the scene it generates3.