Modern chatbots based on large language models - ChatGPT, Claude, Google Gemini, Perplexity AI - imitate human speech impressively: they write poetry, offer ideas, and solve complex tasks. Their creators were inspired by the work of the brain and built artificial neural networks in the image of living neurons. However, similarity of results does not mean similarity of processes.
A person reaching for a cup at the edge of a table instantly calculates the consequences: a wrong movement and it will break. Such foresight relies on expectations that we have accumulated through interacting with the world. One influential theory - predictive processing - suggests that perception is partly structured this way: the brain builds expectations, compares them with incoming information, and corrects them when they do not match. If you pick up an empty cup thinking it is full, your hand applies extra force - and feedback helps refine the movement.
Robots need long training and simulations to manipulate objects reliably. Language models also predict, but something different: the next word or fragment of text. From the phrase "A hand knocks a cup at the edge of the table. The cup..." a chatbot will with high probability continue: "...falls to the floor." It does this based on patterns in huge data sets, without ever having held a cup in its hands. The ability to describe a plausible outcome does not yet indicate an understanding of causes.
This is where "world models" come in - a class of approaches in AI that learn to predict how situations develop and, in some cases, how actions change the course of events. They can learn from video, simulations, or recordings from a robot's sensors. For a home robot, this means, for example, foreseeing that too strong a grasp will break the cup. Preliminary studies show that if actions, movement, and sensory information are taken into account, such systems help link predictions to the physical world.
But an identical prediction does not reveal what knowledge stands behind it or how it was obtained. For a person, a cup is not just a cylinder on a table, but an object that can be picked up, broken, that belongs to someone, is hot, or is about to fall. Our understanding links physical properties to a network of possible actions and consequences. And it forms early: infants already have basic expectations about the behavior of objects at a few months old. Psychologist Jean Piaget believed that children build knowledge through action, and later learn to mentally foresee the result without performing the action.
AI with world models can also learn from actions and sensory feedback - just like people. The question is how its experience turns into usable knowledge. AI may well approach foresight and use prediction for action, but this in itself does not prove that it perceives, learns, or understands the world the same way we do. The better AI reproduces the results of human thinking, the more important it is to distinguish performance from process.