Source: French to English Tester Published on: 2026-05-06
Source: The Conversation – in French– By Julien Perez, Associate Professor – AI and machine learning, EPITA
Generative artificial intelligence systems, which speak so well, do not yet understand the world. New physical or statistical methods such asworld models, or “world models,” would enable them to have a form of common sense, which would help them better simulate reality and interact with it more effectively.
Imagine a child who, after seeing a ball roll behind a couch, instinctively knows that it continues to exist and can anticipate the exact place where it will reappear. This fundamental ability, which psychology calls thepermanence of the object, constitutes a foundation of human intelligence. We do not merely react to the images that strike our retina; we constantly simulate the future in our mind.
Today, artificial intelligence is trying to cross this decisive threshold. After the era of models capable of generating text, like ChatGPT, or images, like Midjourney, a new frontier is emerging with theworld models(“world models”). The challenge is significant: it is about equipping machines with a form of physical, spatial, and logical common sense so that they stop imitating… and finally begin to understand.
Also to read:
Is the mind a predictive machine? Introduction to Bayesian brain theory
These models are already showing promising results in the laboratory or in simulated environments. However, their maturity remains limited and their actual deployment is still restricted today.
Why are current AIs still partly limited?
The most famous AI systems today aregenerative models, like Claude or ChatGPT. These excel at predicting the next word in a sentence or the next pixel in an image, relying on monumental statistical correlations.
From this basic idea, thefirst measurable evidence of reasoning and functional common sensehave been observed in the history of artificial intelligence (AI). However, as regularly pointed out by researchers in the field, such asYann LeCun, Scientific Director of AMI Labsor Fei-Fei Li,Scientific director of Worldlabs, these models do not have a coherent internal representation of physical reality.
That is what notably explains their famoushallucinationsAI: a language model can assert with full confidence that a cow egg is a classic cooking ingredient, simply because it manipulates concepts without fully understanding the biological constraints of the real world. To go beyond this stage of «stochastic parrot” (“stochastic” referring to a phenomenon or model that incorporates randomness in a structured way, like a probability calculation where the unexpected becomes a key data point), AI must incorporate an architecture capable of modeling causes and effects.
This ambition is not new, but it today benefits from an unprecedented technological alignment. As early as 1943, the neuroscientistKenneth Craik already suggestedthat the human brain works by building small-scale models of reality to anticipate events. Thus, when crossing the street, our brain imagines in advance the trajectory of the cars to know when it is safe to cross.
What has changed since then is that we now have enough computational power and mathematical frameworks to test this hypothesis at the scale of complex machines. Interest in these models notably exploded after theworks pioneersby David Ha and Jürgen Schmidhuber, in 2018. They showed that an AI could learn to drive in a virtual environment by training almost exclusively in its own “dreams.” These “dreams” correspond to an internal simulation, created by the AI itself, which allows it to test different strategies without interacting with the real world.
The architecture of world models
These authors introduced the notion of a “world model”: an internal and structured representation of an environment that allows an agent to anticipate the consequences of its actions. Thevirtual modelsynthesizes observable information to construct an abstract and manipulable version of the real world, facilitating planning, simulation, and decision-making, even in complex or uncertain situations. Technically, a world model relies on a mechanism of information compression and prediction.
Rather than merely identifying objects as “cat” or “ball” after learning, a world model learns to represent the world in a richer and more structured way.
At first, the system observes enormous quantities of data and extracts one from itcompact representation of essential dynamics, for example the trajectory of an object, the rigidity of a surface, or the spatial interactions between several elements (the cat’s paw playing with the ball). This abstraction is not limited to labels: it captures physical and logical regularities of the world.
Secondly, the model can simulate future scenarios using this representation (the ball passes under an armchair and the cat tries to clear it). Thus, if the agent equipped with the previously described world model considers an action, it can predict its consequences even before executing it, in a potentially uncertain or noisy environment.
In other words, unlike the simple statistical classification “this is a cat,” the world model learns a kind of internal mini-simulation of the world, which combines perception, spatial understanding and logic, and the ability to anticipate.
Here, the approach remains statistical, similar to reinforcement learning, but without direct recourse to explicit physical models; it is based solely on the regularities observed in the data (balls rolling under objects either come out or remain stuck). This distinction between statistical and physical approaches becomes important when dealing with complex and uncertain environments, where predictions must account for the natural variability of the real world.
Also to read:
Paths for trustworthy AI: blending human expertise and machine learning
Several recent proposals illustrate the potential of the statistical approach to world models. The modelMeta’s V-JEPAlearns, for example, to understand complex physical interactions simply by watching videos, without any human labeling. On its side, Google DeepMind recently unveiledEngineering, an architecture capable of creating interactive virtual worlds from a simple photograph, proving that the machine haspreviously assimilated the laws of physics and perspective.
Applications that affect society
The repercussions of this technology are massive and go far beyond the realm of theoretical computer science.
In robotics, for example, aagent equipped with a world modelcould learn tohandle fragile objectsor to move around in a cluttered warehouse without going through thousands of costly and risky hours of physical trials.
In the autonomous vehicle sector, pioneers, such asWayve, claim to useworld modelsso that cars anticipate the difficult-to-predict behaviors of pedestrians or other drivers, where conventional systems would only react with a delay.
In the field of health, thedigital twins are still in the exploration phaseand are used to simulate how a disease might evolve in response to an experimental treatment. However, these models do not provide certain predictions: they are said to be “probabilistic,” which means they are based on probability calculations. In other words, they estimate several possible outcomes for a patient (improvement, stability, worsening) and assign each a likelihood of occurring, based on the available data and statistical models. Consequently, these simulations remain estimates, not certainties. They must therefore be validated with great rigor, especially when they concern treatments that have never before been tested under real clinical conditions.
The progress of AI leads us to rethink what it really means to “understand” and “anticipate” in a complex world. In the long term, exploring these questions could not only transform technology but also our way of apprehending human cognition and creativity.
It is important to temper the enthusiasm around these models. Despite the advances, they currently remain at the scale of research and development. For example, in robotics and autonomous vehicles, the majority of applications are still in the prototype or controlled pilot stage, often in very structured environments.
Large-scale adoption will require overcoming technical challenges andmajor regulatory, such as robustness in the face of unforeseen situations or safety in complex real-world contexts. Thus, these models are in an advanced experimentation phase, and not operational everywhere and all the time — even if their prospects remain very promising.
Julien Perez is a member of bpifrance, director of AI research.
–ref. The “world models”, when artificial intelligence learns to understand the world –https://theconversation.com/world-models-when-artificial-intelligence-learns-to-understand-the-world-281055