Large language models (LLMs) are increasingly proposed for construction-site safety question answering, yet the evidence base remains small, recent, and methodologically mixed. This review synthesises 17 studies identified through a PRISMA-guided search of Scopus and Web of Science, supplemented by backward snowballing and structured quality appraisal. Unlike broader surveys of artificial intelligence in construction, the paper focuses on how LLM-based safety QA systems are adapted for reliability in high-stakes site contexts. The corpus is organised into three intervention layers: prompt and interaction design, retrieval and knowledge grounding, and model-level architectural adaptation. Prompt-based approaches are accessible and common, but remain constrained by prompt sensitivity and limited source traceability. Retrieval-augmented and knowledge-graph approaches offer stronger support for factual accuracy on regulation-intensive tasks, although they depend heavily on curated repositories. Model-level adaptation offers deeper domain specificity for bounded tasks, but faces data, cost, and maintenance barriers. VOSviewer maps further suggest a shift from generic ChatGPT experimentation towards grounded, domain-adapted and multimodal systems.
Nguyen et al. (Mon,) studied this question.