Can AI truly understand cause and effect? New research reveals that even the most advanced large language models (LLMs) like GPT-4 struggle to reason about causality in narratives, often relying on unreliable shortcuts. Researchers at Carnegie Mellon University explored how LLMs infer causal relationships from stories, uncovering some surprising weaknesses. They found that LLMs often default to assuming that events mentioned earlier in a story cause later events, regardless of the actual causal relationships. This 'topological ordering' bias means that simply changing the order of events in a narrative can drastically alter the AI's understanding of causality. Furthermore, LLMs often prioritize their pre-existing knowledge, or 'parametric knowledge,' over the information presented in a specific story. Even when a narrative explicitly contradicts common sense, the AI might stick to its internal biases, ignoring crucial contextual details. This over-reliance on pre-programmed knowledge hampers the LLM's ability to adapt to new situations and learn from novel scenarios. The study also showed that LLMs struggle with longer narratives, suggesting limitations in their ability to track long chains of cause and effect. However, there's a glimmer of hope. Researchers discovered that when prompted to explicitly construct a causal graph representing the relationships within a narrative, the LLM's performance significantly improved. This suggests that guiding the AI to focus on the overall structure of events, rather than relying on surface-level cues like word order, can unlock its potential for causal reasoning. The implications of this research are far-reaching. As LLMs are increasingly used in areas requiring complex reasoning, like medical diagnosis or legal analysis, addressing these failure modes becomes critical. Developing techniques to enhance causal reasoning in AI is essential for building truly intelligent and trustworthy systems.