Can AI actually drive scientific discovery?

Language models are great at processing text, but they struggle with true innovation. Researchers at Google Deepmind suggest these tools lack the mental models required to discover brand new scientific facts. They are essentially just predicting patterns instead of reasoning.

To move forward, we may need a different approach. The experts believe we need world models that understand how physical systems actually function. These systems would need to observe and interact with the world to build real knowledge.

True scientific revolutions require more than just pattern matching. We need AI that can test hypotheses and learn from experience. Until then, these tools will remain smart assistants rather than independent scientists.