What is a 'world model', and why some AI researchers see it as an alternative to LLMs

Ask most people to picture artificial intelligence today and they will describe something like a chatbot: a large language model, trained on enormous quantities of text, that predicts the next word in a sequence with uncanny fluency. But a growing number of AI researchers argue that a different kind of system, known as a world model, is better suited to tasks that require understanding how situations unfold over time rather than simply generating plausible-sounding text.

A world model, broadly defined, is a system trained to build an internal representation of how an environment works and to predict how it will change in response to actions or events. Rather than learning primarily from text, world models are often trained on sequences of observations, such as video, sensor data, or structured events, with the goal of anticipating what happens next rather than what word comes next.

The distinction matters because language models and world models are optimized for different things. A large language model is fundamentally a text-prediction engine; its fluency comes from statistical patterns learned across vast bodies of written material. A world model is trying to capture something closer to cause and effect: if this happens, what is likely to happen next, and how does that unfold across time.

Alex LeBrun, executive chairman of the AI medical scribe company Nabla, has become one of the voices arguing that this distinction is not just academic. Speaking to STAT News, LeBrun laid out why he believes world models may ultimately prove more useful than large language models for certain categories of real-world tasks, medical documentation among them.

The argument centers on the nature of clinical encounters. A doctor's conversation with a patient is not a static block of text to be summarized; it is a dynamic, evolving interaction in which context accumulates, symptoms are described out of order, and meaning depends heavily on what happened earlier in the exchange. Proponents of world models argue that systems built to track how situations evolve, rather than simply pattern-match on language, may handle that kind of complexity more reliably.

This is not a new idea in artificial intelligence research broadly. World models have roots in reinforcement learning and robotics, where systems have long been trained to predict the physical consequences of actions in order to plan effectively. Video-game-playing AI agents and self-driving car systems have used variations of the concept for years, often described as learning a simulation of their environment that they can use to plan ahead.

What has changed recently is the growing application of world-model thinking to domains once assumed to be the exclusive territory of language models, including fields like healthcare documentation, customer interactions, and other structured but conversational tasks. Some AI labs have begun publishing research explicitly framing world models as a complementary, or even eventual successor, approach to today's dominant large language model paradigm.

Skeptics caution against overstating the divide. Many modern AI systems already blend elements of both approaches, using language models for fluent output while incorporating structured reasoning or memory components that resemble aspects of world modeling. The two approaches are not necessarily mutually exclusive, and some of the most capable systems being built today combine them rather than choosing one over the other.

For an industry still working out where large language models excel and where they fall short, the world-model argument offers a useful frame for thinking about limitations. Tasks that require deep tracking of unfolding, real-world context, rather than generating a plausible next sentence, may simply require a different kind of underlying architecture, whatever it ends up being called.

Whether world models become a mainstream complement to large language models, or remain a research niche championed by figures like LeBrun, will likely depend on whether they can demonstrate clear, measurable advantages in production settings such as clinical documentation. For now, the concept offers a useful vocabulary for a debate that is likely to intensify as AI companies look beyond chatbots toward systems meant to operate reliably in the messiness of the real world.