World models for lawyers

As world models move from research into robotics and autonomous systems, lawyers will confront new questions about auditing, liability and AI governance.

Contributors:
Rafal Fryc
Incoming Associate
Hogan Lovells Cadwalader
World models are a type of artificial intelligence that learns how the physical world works from video and sensor data to simulate what will happen next. This enables robots, autonomous vehicles and other physical systems to plan and act in their environment and will could help power the future of autonomous robotics.Â
As these models are deployed, lawyers will face a new kind of challenge: How do organizations audit, certify and assign liability for AI systems with their own understanding of reality?
World models predict the next state of reality
A video model answers, "Given these words, what image comes next?" A world model answers, "Given the current state of the world and an action, what happens next?"
World models operate by encoding a wide variety of inputs from different sensors and video into a compressed representation of the world's physical dynamics. A video model knows that fire and smoke co-occur in images. A world model knows that fire causes smoke, that heat rises, that materials combust at specific temperatures, allowing it to simulate what happens next in a novel scenario.
The key distinction between video models and world models is interactivity. Video models generate a scene in one shot from a prompt; the output is a fixed clip that is finished the moment it renders. World models predict the next scene based on an action taken by an agent or user, producing frames step by step as each new action arrives. The output is responsive rather than fixed, creating something closer to a video game than a film.Â
Two kinds of world modelsÂ
Generative world models simultaneously produce video frames that a human can observe and an interactive environment that a computer can act within. That responsiveness makes the simulation useful for training policies, which are essentially agents trained to make decisions that control physical outputs. The playable simulation can expose a policy to countless scenarios that are not feasible to achieve in the real world. The result is a policy that can be deployed inside a physical machine to control actions based on the external input it receives, with all of its training done in a virtual world made to resemble ours.Â
Latent world models predict future states entirely in an abstract mathematical space, known as latent space, and never produce visible output. Unlike generative world models which create an environment to train policies, latent models themselves can be put into machines, essentially integrating the simulation and the decision-maker. For example, Meta's V-JEPA 2 compresses object positions, velocities, and physical relationships into latent representations and plans entirely in that compressed space, simulating a robot arm's movement without generating a single image and testing which candidate action best matches the goal, in milliseconds.Â
Both approaches have attracted serious capital and intellectual backing. Meta is betting that latent-space prediction is a path to genuine physical understanding. Meanwhile, Wayve, General Intuition and Decart, who are taking the generative path, have attracted billions in funding.
Legal implications
Both generative and latent world models raise the same fundamental legal question: How can organizations demonstrate that AI systems acting on their understanding of the physical world are safe and reliable? What happens when a world-model powered robot gets its "Mythos" moment? After U.S. President Donald Trump's administration deemed Anthropic's Mythos/Fable model to be too dangerous to release, Anthropic had to undergo evaluations and demonstrate the effectiveness of their model's safeguards in order to re-release it to the general public. If a world model gets deployed in a high-risk environment, it will likely meet a similar level of scrutiny.Â
Auditing will be a key component of world model safety demonstrations, with some parts of robotics law already filling the gaps. For industrial robots, there are task-based risk assessments recommended by the U.S. Occupational Safety and Health Administration. In medicine, the FDA's intended-use doctrine would govern. As a catch-all, there is also product liability in court. Each of these areas would require some kind of assurance that a robot will act as intended.Â
Generative world models are partially auditable at the output level since a human auditor can watch the trained policy behave through video and spot obvious physics errors. The reasons for the prediction stay buried in the model's internal representations, but the prediction itself is represented in visual form, offering some understanding of the virtual world. However, even this is limited. No checklist catches the physics error that the auditor doesn't notice. The auditor ultimately relies on trained judgment about whether the digital world behaves as the real one would, similar to the airplane pilots that test flight simulators to satisfy the subjective evaluation for simulator certification.Â
Latent world models don't natively produce video output that can be inspected. It is possible to build a decoder, separate from the world model, that maps the latent space back to pixels. But a decoder is not the world model displaying its prediction. It is a second model approximating the first. The latent model's prediction of physics is an approximation of real-world physics, and the decoder's rendering of the latent space is its own learned approximation. Rather than testing a flight simulator by sitting in the cockpit, an auditor would basically have to play a video game whose pilot character sits in an in-game flight simulator. Here, we have two layers of rendering between the auditor and the system under review and no way to verify that either layer is faithful to what would happen in the real world.
There are three areas that lawyers should pay attention to:
Liability allocation across the world model supply chain is undefined
This is already an existing problem in AI governance: there is no defined liability allocation for product defects in world models. Called the foundry problem in a paper by Stanford's CodeX, it is unclear whether liability should rest with the foundational model provider, deployer or operator when the model fails. Unlike a chatbot's wrong answer, a world model failure could mean a vehicle that brakes late or a robot arm that misjudges its grip strength.
There is no established method to certify that a world model matches the real world
Both generative and latent world models are learned approximations of physics. Existing high-risk industries such as finance, aviation and nuclear, have agreed statistical evidence standards and reference datasets for compliance, while world models lack equivalent benchmarks. Early work exists, from DeepMind's Physics-IQ benchmark, where photorealistic video models score dismally on physical principles, to the U.S. National Institute of Standards and Technology's measurement-science program for physical AI, but nothing approaches a certification standard.Â
Regulatory fragmentation creates compliance complexityÂ
In the U.S., there is not a federal AI governance law, which enables more than 50 jurisdictions to issue their own binding law on any emerging technology. There have been some initial attempts at legislation for robotics, and it is likely to continue to grow. However, if world model/robotics legislation mimics other emerging technologies, such as chatbots, then we are likely to see a heavily fragmented approach with dozens of states creating their own definitions, requirements and penalties. Fragmentation would then make auditing world models more difficult to implement, as each jurisdiction could set a different standard to audit against.Â
Most of the current efforts at governing AI, such as the voluntary NIST guidance, lack specific requirements for physical AI and world models. World models challenge many of the assumptions that legal frameworks for evaluations, supply-chain liability, and safety auditing all depend on. The topic is ripe for exploration.
Questions to consider:
- What would meaningful auditing look like for world models?
- Can existing AI frameworks adequately cover world models, or do they require their own regulatory approach?
- How should we legally define a false understanding of physics?Â
The author would like to thank IAPP Managing Director, Washington, D.C., Cobun Zweifel-Keegan, CIPP/US, CIPM, and Hogan Lovells Cadwalader Senior Associate, Data, Privacy and Cybersecurity, Julian Flamant, CIPP/E, for their thoughts and guidance.
This content is eligible for Continuing Professional Education credits. Please self-submit according to CPE policy guidelines.
Word count: 1319

Contributors:
Rafal Fryc
Incoming Associate
Hogan Lovells Cadwalader



