World models: from generation to simulation

The market conversation around AI has moved quickly: from generative AI, designed to produce or transform content across text, image, audio and video, to agentic AI, in which systems plan, decide and act within a workflow. One of the next categories attracting attention is world models.

The market signals are notable. In February 2026, Fei-Fei Li’s World Labs announced¹ $1 billion in new funding for its work on spatial intelligence and world models. In March 2026, Yann LeCun’s Advanced Machine Intelligence announced² a $1.03 billion raise to build AI systems focused on reasoning, planning and world models. 

Those announcements suggest that major investors see world models as a serious area of development.

What are world models?³

Generative AI models are typically associated with producing outputs: text, images, audio or video. World models are aimed at something different. Their purpose is to model how an environment works, how it changes over time, and how actions may affect it.

For example, a video generation model asked to show a robot picking up a cup may produce something visually convincing. A world model is trying to capture something deeper: what happens when the robot moves, how the cup shifts, what resistance the robot encounters, and what happens if the grip is slightly wrong. The focus shifts from appearance to cause and effect.

Put simply, a world model attempts to represent an environment in a way that allows future states to be predicted. In some cases, the result may be a human-readable simulated world. In others, it may be an abstract representation used by an autonomous system to plan, test or evaluate possible actions.

Fei-Fei Li and World Labs have described this broader direction as part of the push towards “spatial intelligence”⁴: AI that can perceive, reason about and interact with the three-dimensional world, rather than simply generate representations of it.

Why they matter

The importance of world models lies in what simulation may make possible.

If a simulated environment captures relevant real-world conditions with sufficient fidelity, a system trained or tested against it may perform more reliably when deployed. That is potentially significant for agentic AI. In theory, an agent operating within or alongside a world model may be able to test action sequences, compare likely consequences and refine its approach before acting outside the simulation.

That is why leading labs connect world models to robotics, spatial computing and physical AI, rather than treating them only as another form of content generation. The commercial promise is not just better synthetic content, but better systems for modelling environments, testing decisions and supporting action.

The legal significance

The legal significance lies not only in the action eventually taken, but in the simulation layer used to support or justify it.

If developers and deployers rely on simulated environments to make claims about safety, robustness or performance, the simulation itself may become part of the governance and assurance perimeter. The question is not simply whether an AI system produced a defective output. It is also whether the simulation-based training, testing or planning process was reasonably designed, validated, documented and relied upon.

That raises practical questions for legal and compliance teams. What assumptions were built into the simulated environment? How closely did it reflect the real-world conditions in which the system would be deployed? What edge cases were tested? What limitations were known? What records exist of training, testing and evaluation inside the simulation? Were the simulated conditions representative, and were the results capable of independent review?

Those questions matter because simulation can create a false sense of confidence if the modelled environment is too narrow, too clean or too detached from deployment reality. A system may perform well in simulation and still fail when exposed to real-world variation.

Looking ahead

Most legal and regulatory discussion of AI remains centred on LLMs and the agentic systems built on top of them. World models point in a different technical direction: one focused less on generating plausible outputs as an end, and more on modelling environments well enough to support action within them.

For lawyers advising on AI, world models are worth tracking now. If the technology moves from research and demonstration into applied settings, the legal questions will not only concern what the AI system did, but also the modelled world in which it learned, planned or tested what to do.