Opinion: How Can Causal AI Benefit World Models?

World models are regarded as a key pathway toward achieving general artificial intelligence, yet current modeling approaches suffer from correlation limitations that hinder their ability to capture the intrinsic causal mechanisms. This deficiency results in significant shortcomings in out-of-distribution generalization capabilities, sample efficiency, and deep reasoning abilities of world models. This paper argues that integrating principles from causal science is essential for overcoming these challenges and constructing world models aligned with core objectives. We systematically propose a framework where the three pillars of causal science address these shortcomings. Ultimately, we contend that the shift from a correlation-driven paradigm to a causality-driven paradigm represents not merely a technical refinement, but a necessary leap toward constructing agents that genuinely understand and interact with the real world.