Imagine an AI that could read any how-to guide, from fixing a leaky faucet to baking a soufflé, and instantly understand the steps involved. This dream is closer than you think, thanks to advancements in Large Language Models (LLMs). However, a new research paper, "Human Evaluation of Procedural Knowledge Graph Extraction from Text with Large Language Models," reveals that while LLMs show promise, there are significant hurdles to overcome. The researchers explored how well LLMs can extract procedural knowledge – the "know-how" of completing a task – and transform it into a structured, machine-readable format called a Knowledge Graph. Their method involved crafting clever prompts, essentially instructions given to the LLM, to guide the AI to identify the steps, actions, objects, and even the time required for each part of the process. They then asked human participants to evaluate the AI’s performance. Interestingly, humans themselves often disagreed on the best way to structure a how-to guide, revealing the inherent subjectivity of this task. While the evaluators generally agreed that the AI accurately captured the essence of the procedures, they were less convinced of its *usefulness* – its ability to present the information in a truly helpful way. This suggests that while AI can extract the *what* of a procedure, it struggles with the *how*. There’s also a hint of human bias at play: participants were slightly harsher on the LLM compared to hypothetical human annotators. This fascinating study illuminates the path toward a future where AI can truly understand and assist us with everyday tasks. But it also highlights the importance of human feedback and the challenge of making AI not just smart, but genuinely helpful.