Abstract: Judea Pearl’s three-tiered ladder of causation was once a widely accepted critique of AI and a guiding program for AI practice. However, with the emergence of some phenomenal large language models such as GPT-4, its AI critique has been thoroughly disproved, and the causal theory of Structural Equation Modeling behind it has also been faced with challenges. The goal of this paper is threefold: first, it intends to elaborate on the core elements of the causal triad and the inner mechanisms of the large language Model, and to show the sense in which the latter reinvents Pearl’s original conception. Second, it also clarifies the theoretical significance for structural causal models of the emergence of such a large-language model capable of “demonstrating causal competence,” which destroys Pearl’s critique but opens up new possibilities for causal research. Finally, and more importantly, the theoretical value of structural causal models remains, and the idea of equipping intelligences with causal inference engines to help them make causal inferences is not obsolete.

Key Words: Structural causal models; Large language models; Ladder of causation