Imagine a security camera that not only records but also understands what it sees, flagging unusual events without constant human monitoring. That's the promise of video anomaly detection (VAD), a field of artificial intelligence focused on identifying unexpected or rare occurrences in video footage. Traditional VAD methods, while effective, often work like a black box, simply providing an anomaly score without explaining *why* something is flagged as unusual. This lack of transparency can limit trust, especially in critical applications like autonomous driving or security surveillance. A new research paper, "Follow the Rules: Reasoning for Video Anomaly Detection with Large Language Models," introduces an innovative approach called AnomalyRuler, which aims to make VAD more transparent and adaptable. AnomalyRuler leverages the reasoning power of large language models (LLMs) to detect anomalies by learning and applying rules about normal behavior. This process mimics how humans learn to spot unusual activity – by first understanding typical activities, then flagging any significant deviations. AnomalyRuler works in two stages: induction and deduction. During induction, the system is shown a few examples of normal video frames and generates descriptive text, from which the LLM extracts rules about typical activities and objects. In the deduction stage, AnomalyRuler analyzes new video frames, also converting them into text descriptions. These descriptions are compared to the learned rules to identify potential anomalies. The system doesn’t need extensive training data of anomalous events, making it adaptable to various scenarios. The researchers demonstrated AnomalyRuler's effectiveness on four benchmark datasets, achieving state-of-the-art performance while also providing clear explanations for its detections. For example, if the system learned that "walking" is a normal activity in a specific area, it would flag someone "riding a bicycle" as anomalous because it deviates from the established norm. This rule-based reasoning makes the system's decisions transparent and understandable, increasing confidence in its reliability. While the research shows promising results, challenges remain. The effectiveness of AnomalyRuler depends on the accuracy of the visual and language models it employs. Future research could explore enhancing the robustness of these models and adapting the system to even more complex and dynamic environments. This work opens exciting possibilities for trustworthy and explainable AI in video anomaly detection, paving the way for broader real-world applications.