OpenAI's Sora has been widely described as a "world model," but Stanford researchers just challenged that claim with a rigorous framework showing it's actually just a sophisticated visual renderer. On June 3, 2026, the World Labs team, in collaboration with Stanford University Professor Fei-Fei Li, published a conceptual analysis titled "A Functional Taxonomy of World Models," arguing that the term "world model" has become one of the most misused phrases in artificial intelligence today.

What Exactly Is a World Model, and Why Does It Matter?

The confusion started in February 2024 when OpenAI released Sora with a technical report boldly titled "Video Generation Models as World Simulators." At the time, Jim Fan, Director of Robotics at NVIDIA, commented on LinkedIn that Sora was essentially a "world model that allows no action as the only action." Since then, the term has been applied to everything from Tesla's autonomous driving prediction systems to game engines to robotic control models, creating a conceptual mess.

Jim Fan, Director of Robotics at NVIDIA

The problem is that these systems do fundamentally different things. A video generator cares about pixel fidelity. An autonomous driving system cares about predicting precise 3D positions and velocities of road participants. A robotics company cares about whether pushing a cup 5 centimeters to the left will cause it to tip over. They're all called "world models," but they're not doing the same thing at all.

How Does the New Framework Categorize AI Systems?

Fei-Fei Li's team returned to a foundational theory from the 1960s called partially observable Markov decision processes (POMDP) to create a clearer taxonomy. This framework describes a complete cycle of interaction between an agent and its environment: the agent observes its surroundings, takes an action, the environment changes, and the agent receives new observations that drive the next action.

Within this framework, the researchers identified three distinct functional categories that current AI systems actually occupy:

• Renderers: Generate high-fidelity pixel outputs for human viewing without genuine physical understanding. These systems optimize for visual realism, not accuracy.
• Simulators: Produce precise physical states suitable for subsequent computations, like NVIDIA Omniverse, which can calculate whether a structure will collapse or how liquid will flow.
• Planners: Output actions for embodied agents, such as Vision Language Action (VLA) models used in robotics that decide what a robot should do next.

Sora falls squarely into the renderer category. The World Labs article explicitly states that rendered buildings may appear "unstable" because the system doesn't actually solve structural mechanics equations, and splashes of liquid may look realistic but bear no relation to real-world physical quantities.

Why Can't Sora Be Used for Tasks Requiring Physical Accuracy?

This distinction has real practical implications. Because Sora is a renderer, not a simulator, it cannot be used for architectural design, robot training, or any task requiring physically accurate simulations. The system learns statistical patterns from massive amounts of video data to generate visually plausible scenes where a cup breaks when dropped and a person's legs swing alternately when walking. These scenes appear to "understand physics," but they're just pattern matching on pixels.

The value of World Labs' analytical framework lies in providing a comparative coordinate system that transcends marketing rhetoric. No matter how a company packages its product, placing it back into the POMDP cycle and examining its inputs, outputs, and missing components reveals its true capabilities and limitations.

How Does This Clarification Affect Investment and Development?

The conceptual confusion has lasted for more than two years, driven partly by the grand narrative quality of the term "world model." It sounds more imaginative and better suited to supporting high valuations and fundraising stories than phrases like "video generation model" or "video prediction model." When technical capabilities fail to match public expectations, the concept inevitably degenerates into a marketing tool.

By systematically clarifying what each category of system can and cannot do, Fei-Fei Li's team has provided a framework that helps prevent misinterpretation, guides investment decisions, and sets a baseline for future integration of these technologies. Google's Genie 3, various text-to-video models, and nearly all AI video generation tools fall into the renderer category alongside Sora.

The research underscores a broader shift in how the AI industry is maturing. Rather than releasing a new model or announcing a benchmark, World Labs chose to do something more fundamental: return to theoretical first principles and clarify the conceptual foundations that have become muddled by hype and marketing narratives.