Large language models provide new technical support for smart policing with their capabilities in natural language understanding, text generation, semantic retrieval and task reasoning. They can improve policing efficiency in alarm summarization, document assistance, police knowledge question answering, anti-fraud warning and clue integration. However, their application may also cause risks such as hallucinated outputs, data security vulnerabilities, algorithmic opacity, unclear responsibility and technological dependence. To ensure standardized application, governance mechanisms should be improved in private deployment, data desensitization, police knowledge base construction, human review, hierarchical authorization, log retention and model evaluation.