Generative AI in Construction Safety: A Systematic Literature Review of Methods, Evidence, Deployment Gaps, and Future Directions

Generative artificial intelligence (GenAI) is increasingly being explored for construction safety; however, its maturity and practical reliability remain unclear. This study systematically reviews GenAI applications in construction safety and assesses whether existing evidence supports reliable, source-grounded, human-supervised, and deployable applications. The review followed a predefined protocol and PRISMA 2020. Scopus, Web of Science, IEEE Xplore, and Google Scholar were searched in May 2026 for studies published from 2019 to August 2026. After duplicate removal and screening, sixty-five studies were included. Data on GenAI methods, safety applications, validation, deployment, and limitations were extracted and synthesized thematically. Six application areas were identified: safety knowledge retrieval and decision support; incident report analysis and risk prediction; multimodal hazard reasoning; synthetic data generation; worker-centered training, communication and behavior support; and generative optimization, monitoring, and safety control. Several studies provided proof-of-concept evidence using historical datasets, simulations, laboratory experiments, or expert assessments, while evidence of generalizability, field deployment, hazard-exposure reduction, and incident prevention remained limited. GenAI showed potential to support hazard recognition, decision support, and compliance-related safety tasks, but field validation, source grounding, and integration with preventive controls are needed for reliable deployment. A conceptual framework linking GenAI capabilities, and preventive safety controls is proposed.