Google DeepMind CEO Demis Hassabis has made it clear: today's large language models don't understand reality. In a recent statement, Hassabis argued that these systems are fundamentally limited and proposed a different path forward — one built on 'world models' that prioritize causality over language processing.
Why language models aren't enough
Hassabis said language models cannot grasp the underlying mechanics of the world. They process text, but they don't build a mental model of cause and effect, he argued. That's a problem if AI is ever going to move beyond pattern matching into genuine reasoning. The criticism is notable because it comes from the head of one of the world's leading AI labs, a lab that itself has invested heavily in language models like Gemini.
World models: causality first
World models, as Hassabis describes them, emphasize understanding how things work — not just predicting the next word in a sentence. They try to simulate the physical and logical relationships between objects and events. Instead of learning from terabytes of text, a world model would learn from interactions in simulated or real environments, building a sense of what causes what. That's closer to how humans learn, Hassabis said.
What this shift could mean
If world models become the dominant approach, the implications reach far beyond chatbots. Robotics is one area where causality matters more than syntax. A robot that can plan actions based on a mental model of its surroundings is far more useful than one that merely completes sentences. Scientific discovery is another. Hassabis pointed out that understanding causal relationships is essential for making predictions and designing experiments — tasks where language models currently fail.
A debate that's just beginning
Hassabis's position puts him at odds with much of the current AI industry, where language models are the hottest commodity. But he's not alone in questioning their limits. The push for world models could reshape research priorities inside Google DeepMind and beyond. No timeline has been given for when such models might reach the capabilities of current language systems, but Hassabis made clear that he thinks the direction is worth pursuing.
