World models offer a promising solution in ensuring that AI systems perform well. This blog examines how building and applying effective world models can increase trust, adaptability, and collaboration between AI and people, making them a cornerstone for responsible and reliable artificial intelligence.

Imagine you’re planning a trip to a new city. Before you go, you look at a map, read a travel guide, and maybe ask friends for tips. You build up a mental picture of what the city is like, where the landmarks are, how the subway works, which neighborhoods are safe, what the local customs and etiquette are, and where to find good food. This mental map isn’t just a list of facts; it’s a living model that helps you make decisions, like how to travel around best or how to avoid offending the locals. In the same way, a “world model” for an AI is like its own internal guidebook, built from experience and information. The applied knowledge needed to manipulate its environment effectively, predict what might happen next, and choose the best actions, even if it’s never been in that exact situation before.

Definitions

  • World: the set of relevant, contextual information that an entity operates within.
  • Model: A model is a simplified representation of data that we can use in place of the actual data.
  • World model: A simplified representation of the contextual information an entity operates within.  

World models, conceptually, are cognitive maps that an entity (human, animal, or AI) constructs from sensory input and/or past knowledge to gain an awareness of how its reality works. This causal model of reality allows the entity to align its direct experience with pre-existing knowledge, simulate scenarios, and choose actions based on their predicted outcomes, enabling sophisticated planning and decision-making. The word causal is key here as world models help the entity make sense of information about the current state of its world as well as what led to this state, and how the state will change upon a given action.

World models do not need to have information about the whole world but merely the world in which the entity operates, within its own “bubble of information”. This is the extent to which the world model needs to have information to have the context it needs for the entity to perform. A world model does not need to be complicated to be useful. For example, at the simpler end of the spectrum, a sommelier needs to know everything about wine but nothing more to perform effectively; they only need enough relevant information about their world to be highly effective in that world. At the larger end of the spectrum, an urban planner or engineer would need a much larger world model to be effective in their profession.

World models are not unique to AI. They are a representation of sociological information (the rules we define as society such as a legal system), sensory information (the interpretation, we have as individuals, of the world around us), physical information (physical characteristics of the world). World models allow us to make decisions rooted in a realistic understanding of the context in which the entity operates in a consistent and repeatable way.

While world models may differ between people, the commonalities between people’s world models allow us to collaborate, anticipate, and empathize with each other to solve tasks efficiently. The inverse is also true, when we have contradictory world models, negative consequences can happen such as arguments, misunderstandings and conflict.

Properties of world models in AI systems

Now we have introduced world models and why they are needed, let’s look at some principles of world models to understand them better.

  1. World models are representations of information (including facts, relationships, laws, rules, causality, uncertainty etc.) in a given area, like a “bubble of information”.
  2. World models are a means to an end rather than an end in themselves; we use them to increase performance (such as accuracy, reliability) and trustworthiness of our AI systems.
  3. A world model will contain a variety of information from very general to highly specific, and from well-proven facts to vague hypotheses or beliefs.

These properties ensure that we are adding relevant context to our AI. This is necessary because humans do not trust context unaware AI to act autonomously on their behalf because it is not accurate and reliable enough. Context is added through a world model enabling the AI to operate effectively within a particular world. One example is the use of world models within physical AI in terms of human & AI interactions. Check out this blog from our colleagues at Cambridge Consultants for more information on this topic.

There are also some optional properties that should be considered when building world models such as continuous learning, risk & impact management, and representation of uncertainty. We will investigate these principles as well as many more in future publications.

The application of world models in AI

All types of AI have one basic underlying mission: to look for patterns in data to make predictions. We can make predictions where we have lots of data (interpolate) or where we have limited data (extrapolate). It is often harder to make predictions where we have limited data. World models give us the ability to do this better by allowing AI to hypothesize different scenarios accurately without having directly observed data from that specific scenario. A better understanding of causality in world models is also a powerful mechanism to use knowledge we already have to infer knowledge that we don’t currently have.

At this point you may be wondering how we get a complete and correct world model. The truth is there is no such thing as a perfect model. This is because there may well be missing or conflicting information taken from different inputs. Handling this conflicting information within a world model is very important. It is important to assess and qualify information to be able to have a coherent world model.  

As George Box, a British statistician said, “All models are wrong, but some are useful.” In many cases, it is both useful and desirable for the world model to contain more subjective principles, such as ethical or moral principles. We need to ensure that the system aligns with our moral principles and safety boundaries, even when raw data might suggest harmful actions are optimal. Ethical principles and practices can be added into world models in the same way as laws are to shape the model’s understanding of that world.

The fact that a world model exists is not enough, it needs to be applied by the AI in the correct way. Alignment Engineering is the discipline concerned with deliberate and conscious use of a world model within an overall AI system to align that system to the understanding of the world in a particular world model. We wrote about this recently so do check it out to learn more.

Enabling trust in AI

World models are essential to AI becoming trustworthy, increasing the accuracy and groundedness of the predictions it can make. This increased level of performance allows those who are responsible for the AI’s actions to give more autonomy to that AI because they know that it aligns with their own world model.

Over the millennia of human existence, we have built world models from shared knowledge, guiding how we work and live. This applied knowledge of our world enabled us to become the most effective species on the planet. It is now time that we exploit that great understanding of the world, to ensure AI is aligned with us and works for us, acting on our behalf. This is essential if we are going to have a complex ecosystem of AI everywhere with a shared world model between all those AIs and us.

About the Capgemini AI Futures Lab:

Capgemini’s AI Futures Lab is part of Capgemini’s innovation ecosystem, focusing on exploring the future of AI to prepare the Group and our clients for what is next. The Lab’s activities range from research, thought leadership, internal readiness to client and partner pioneering collaborations. It explores horizon 3 technologies, those on the frontier of AI with transformative potential. The Lab operates like an incubator, rapidly prototyping high-risk, high-reward ideas and charting paths to scale them internally or with clients.