Exploring the Opportunities and Limitations of Large Language Models for Construction Safety Interpretation: Implications for Education and Practice

In the United States, construction and extraction occupations recorded 1,032 fatalities in 2024, sustaining interest in AI tools for safety analysis. Large language models (LLMs) are accessible to construction professionals and students, yet their capabilities and limitations on real safety data remain poorly understood. This study examines how a fine-tuned LLM interprets authentic OSHA Severe Injury Report narratives across three tasks: incident summarization, contributing factor identification, and preventive insight generation. A QLoRA fine-tuned Llama 3.1–8B-Instruct was compared against the base model across 30 unseen evaluation cases, of which 8 served as anchor cases for in-depth analysis. Incident summarization was stable and reliable: all 30 evaluation outputs met the required format, and fine-tuned outputs matched reference responses in 97% of held-out validation cases. Contributing factor identification improved partially, but 33% of fine-tuned factor lines named hazard classes or restated events rather than causal conditions. Preventive insight generation exhibited complete template collapse: all 30 outputs were pre-defined template bullets, with six unique responses, traced to low training-data diversity (25 unique patterns across 269 examples). Findings suggest surface-level output quality does not guarantee analytical depth, and in this setting fine-tuned output quality was shaped more by training-data diversity than by model architecture alone.