The Impact of Low-Quality Accident Reports on the Performance of Construction Safety AI Models

The radical advancement of artificial intelligence (AI) is transforming construction safety management practices. AI models for construction safety are typically trained using accident reports providing detailed information on incidents and contributing factors. However, the models’ performance is compromised by low-quality accident reports containing unclear causal relationships or misclassified information. Given the complex and trade-specific nature of construction accidents, automatically filtering out low-quality reports remains challenging. This study examines how low-quality accident reports affect the performance of AI models classifying construction trades associated with reported accidents. Using 430 accident reports, the study measured model performance after varying proportions of low-quality data were included. The results indicated the performance declined up to 45% in specific classifications. These findings demonstrate that even a few low-quality accident reports can degrade the performance of construction safety AI models, emphasizing the need for advanced filtering mechanisms to ensure the reliability of construction safety AI models.