World models are the most underestimated AI shift of this moment. While boardrooms are still debating ChatGPT, the fundamental architecture of AI is already moving on. The boardroom implications are greater than anything the generative AI hype has ever been.
In early 2025, Meta's research team demonstrated V-JEPA 2 — an AI system that had learned, without explicit instruction, how physical objects behave when they fall, roll, collide or are stacked. The system had never attended a physics lesson. It had not been taught equations. It had watched videos — and from those videos had constructed an internal model of reality that it could use to reason about situations it had never previously encountered.
Around the same time, Google DeepMind presented Genie 2: a system capable of generating interactive three-dimensional worlds from a single image, including consistent physics, causal relationships and navigable space. No hand-coded rules. No pre-programmed laws of the world. Only a learned understanding of how things work.
This is not progress along the same path as GPT-4. This is a fundamentally different architecture — and a fundamentally different form of intelligence.
The large language models that underpinned the AI wave of 2022–2025 — GPT, Claude, Gemini in its early iterations — are exceptionally powerful pattern recognisers. They predict the most probable next word, given all preceding words. They do this with so many parameters, trained on so much text, that the result is impressively coherent. But it is fundamentally statistical: which word fits best here?
A world model does something essentially different. It builds an internal representation of the world — a simulation — that it can use to reason forward about cause and effect, to predict the outcomes of actions before they are executed, and to plan in environments it has never previously encountered.
Yann LeCun — chief AI scientist at Meta and one of the most influential AI researchers in the world — has argued repeatedly that world models are the architecture on which truly autonomous, reliable AI will ultimately be built. His argument: as long as AI is based on language alone, it lacks the causal structure of reality that is necessary for genuine reasoning and reliable action.
A board that treats this as an academic debate is missing the governance urgency. World models are not a laboratory phenomenon of 2030. They are already operational in specific domains — and the sectors where they are being deployed first are precisely the sectors where the greatest value and the greatest risk converge.
Industrial automation and robotics Robots using world models no longer need to be programmed for each specific task in a specific environment. They learn a model of their surroundings and can independently reason about how to execute a new task. Boston Dynamics, Figure AI and 1X Technologies are actively building on this now. The implications for manufacturing, logistics and supply chain are immediate.
Autonomous vehicles and mobility Waymo's fifth-generation autonomous driving software uses world model architecture to anticipate the behaviour of other road users — not by pre-programming every possible scenario, but by building and applying a causal model of traffic behaviour. Reliability is measurably higher than in previous generations.
Scientific research and simulation In the pharmaceutical industry, world models are already being used to simulate the outcomes of molecular interactions — with an accuracy and speed that traditional laboratory research cannot match. The drug discovery timeline compresses by a factor of three to five.
Climate and energy optimisation Google DeepMind's systems for optimising data-centre cooling and electricity grid management use world model principles: they simulate the consequences of actions before executing them. Energy efficiency gains amount to 30 to 40 per cent above what traditional optimisation achieves.
World models are not a revolution arriving without warning. They are a shift already under way — and for boards there are three implications that already demand strategic attention.
The sectors hit earliest by world model-based AI are those with the most predictable physical processes: manufacturing, logistics, energy, transport, pharmaceuticals. If your organisation operates in one of those sectors, the clock is running.
Organisations investing now in world model applications are building a learning advantage that is difficult to close. World models improve the more they interact with the reality they are modelling. Every day a competitor trains its world model on production data, logistics data or patient data, the gap widens.
The competitive advantage that world models confer is not a one-off — it accumulates. A world model that has learned from your production line or logistics network for six months longer is structurally better than one that starts six months later. The board that treats this as "something for later" is creating a catch-up problem that grows exponentially.
The governance discussion of 2023–2025 was largely built around the characteristics of large language models: hallucinations, bias in training data, copyright questions around generated content. These are real risks — but they are the risks of one type of AI architecture.
World models carry a different risk profile. When an AI system plans and executes actions based on an internal model of the world, the possible errors are not "a wrong word in a text" but "a wrong action in physical or economic reality." The scale and irreversibility of errors increases.
Your current AI governance is probably built for LLM risks. It is not yet built for the risks posed by systems that simulate the world and act on the basis of that simulation. That is a governance gap that urgently needs to be filled — not when the systems arrive, but now.
The central question for governance of world model systems is not "is the output correct?" but "is the internal model of the world correct?" A system with an incorrect world model makes systematically wrong decisions — not randomly, but consistently in the direction of its flawed assumptions. That is harder to detect and more dangerous in impact than the hallucinations of an LLM.
Language models are trained on text. The organisations that held the most high-quality text — technology companies, media companies, publishers — had an advantage in the LLM era.
World models are trained on interactions with the world: sensor data, movement data, production data, logistics data, medical imaging data, financial transaction data. The organisations that hold the most rich, structured data about how the world behaves in their domain — manufacturers, hospitals, logistics operators, energy companies — have a structural advantage in the world model era.
This means that data strategy is no longer only about "what data do we hold and how do we secure it" but about "what data enables us to build a superior world model for our domain?" That is a strategic question at board level — not an IT question.
Understand the architectural shift — genuinely Not at the level of a management summary, but at the level of understanding that enables sharp questions. The distinction between pattern recognition and causal reasoning is the most consequential distinction in AI at this moment. A board that understands this asks different questions of its management than one that does not.
Inventory your domain-specific data as a strategic asset What data does your organisation hold that is unique to your domain? Production sensor data, clinical observation data, logistics movement data, financial transaction patterns? That data is the raw material for a world model your competitors cannot replicate. Treat it as such in your strategic planning.
Put the question to your AI advisers Ask explicitly: which of our current or planned AI applications use world model architecture, or could do so? Where in our value chain is causal reasoning — predicting the outcomes of actions — more valuable than pattern recognition? The answers determine where your next investment delivers the highest strategic return.
Update your governance frameworks proactively Begin now thinking about governance for systems that plan and act on the basis of an internal world model. Which decisions may such systems take autonomously? How do you verify the quality of the internal model? What are the stop criteria when the model fails? These questions are theoretical today — in 24 months they will be operationally urgent.
The most used metaphor for AI in recent years has been the calculator: a tool that helps people do faster what they were already doing. World models fit a different metaphor: the junior employee who not only executes what is asked, but also thinks about how something can best be done, anticipates obstacles and independently makes plans.
That is not the same technology at higher speed. That is a qualitatively different form of AI — with qualitatively different implications for organisations, for work and for governance.
The board that enters 2026 with only an LLM strategy is preparing for yesterday's war. The board that understands the world model shift and positions itself strategically is ahead of a development that will rewrite the rules over the next five years in every sector that involves physical processes, causality and planning.
The question Praesum puts to every board in 2026: does your AI strategy understand the difference between a system that predicts words and a system that simulates the world? If the answer is no — that is the first gap to close. Not because world models will take over your market tomorrow. But because the decisions you make now about data, governance and investment will determine whether you are ready when they do.
Not only understanding what world models are — but knowing what they mean for your organisation.