Large Language Models (LLMs) have demonstrated remarkable abilities in various tasks, but can they truly understand cause and effect? Recent research reveals that inferring causal relationships from correlation, a cornerstone of human reasoning, poses a significant challenge for these powerful AI systems. While LLMs can easily parrot back correlations they've encountered in their vast training data, figuring out *why* things happen—distinguishing cause from mere correlation—is a different story. This problem is known as Natural Language Causal Discovery (NL-CD). Imagine an LLM being told that ice cream sales and shark attacks both increase in the summer. A human quickly recognizes a confounding factor—warm weather—driving both trends. However, without this deeper understanding of causality, an LLM might incorrectly conclude that ice cream sales *cause* shark attacks. This inability to reason causally has significant implications for AI applications in areas requiring nuanced decision-making, scientific discovery, and policy analysis. To address this limitation, researchers are exploring innovative prompting strategies. One promising approach leverages the PC algorithm, a well-established method for causal discovery. This strategy, known as PC-SUBQ, guides the LLM through a step-by-step reasoning process, mimicking how a human might apply causal inference rules. By breaking down the complex task of NL-CD into smaller, more manageable sub-questions, PC-SUBQ helps the LLM navigate the intricate relationships between variables and infer causal structures. Initial experiments with PC-SUBQ across several popular LLMs, including Gemini and GPT models, show promising results. This method consistently outperforms simpler prompting techniques, suggesting that guiding LLMs through structured reasoning is crucial for unlocking their causal inference potential. Furthermore, PC-SUBQ exhibits robustness to variations in wording and variable names, indicating a more generalizable understanding of causal principles compared to models trained by simply memorizing examples. This research offers a glimpse into the ongoing journey towards imbuing LLMs with more human-like reasoning abilities. While challenges remain, particularly with increasing problem complexity, the progress demonstrated by PC-SUBQ offers a compelling path towards LLMs capable of not just observing correlations, but truly understanding cause and effect.