science & techanalysisThe Conversation
Why AI cannot ‘think’ like humans do
Large language models learn to predict text, but a newer approach called "world models" aims to give AI an internal simulation of how things work, so a system can anticipate what will happen and act accordingly. For example, a robot could imagine what happens if it pushes a cup before it does so. This is useful for planning and for physical tasks. But predicting outcomes isn't the same as human thinking, which involves lived experience, a body, emotion and an understanding of meaning. A world model is a statistical map built from data, and it can be wrong in ways a person's common sense usually would not be.
A world model is trained on video, sensor data or simulations to predict what comes next in a scene. The system can then try out actions in its internal model, choosing the one with the best predicted result.
The distinction matters for how far to trust AI and what to expect from it. It also shows where research is heading, beyond chatbots, towards systems that act in the physical world.
Researchers such as Yann LeCun have argued that language models alone will not reach human-level intelligence and that world models are needed.
Illustrative photo · JPxG, Public domain, via Wikimedia Commons
