Publisher
University of Tennessee at Chattanooga
Place of Publication
Chattanooga (Tenn.)
Abstract
Occupational injuries create substantial human and financial costs, making effective hazard reporting essential for prevention. This study examines whether large language models (LLMs) can improve the evaluation of Behavior-Based Safety reports and predict future workplace incidents. Building on prior research that used an LLM to score over 33,000 refinery safety reports across ten quality indicators, we will compare multiple LLMs available through Azure AI Foundry using identical reports, scoring rubrics, prompts, and analytic procedures. Models will be evaluated for rating consistency, agreement across models, and their ability to predict next-day incidents through time-series logistic regression analyses. We expect some LLMs to produce more reliable quality ratings and stronger injury-risk predictions than others. Findings will help identify the most effective models for analyzing open-text safety reports and supporting proactive injury prevention. Results may inform the development of automated safety-reporting tools and conversational systems that encourage higher-quality reporting across industries.
Date
November 2026
Subject
Industrial and organizational psychology
Document Type
posters
Language
English
Rights
http://rightsstatements.org/vocab/InC/1.0/
License
http://creativecommons.org/licenses/by/4.0/
Included in
From Words to Warnings: Using LLMs to Predict Injury Risk
Occupational injuries create substantial human and financial costs, making effective hazard reporting essential for prevention. This study examines whether large language models (LLMs) can improve the evaluation of Behavior-Based Safety reports and predict future workplace incidents. Building on prior research that used an LLM to score over 33,000 refinery safety reports across ten quality indicators, we will compare multiple LLMs available through Azure AI Foundry using identical reports, scoring rubrics, prompts, and analytic procedures. Models will be evaluated for rating consistency, agreement across models, and their ability to predict next-day incidents through time-series logistic regression analyses. We expect some LLMs to produce more reliable quality ratings and stronger injury-risk predictions than others. Findings will help identify the most effective models for analyzing open-text safety reports and supporting proactive injury prevention. Results may inform the development of automated safety-reporting tools and conversational systems that encourage higher-quality reporting across industries.
Department
University of Tennessee at Chattanooga. Dept. of Psychology