Abstract
Background/Objectives: The integration of artificial intelligence (AI) into forensic science is expanding, yet its application in firearm injury diagnostics remains underexplored. This study investigates the diagnostic capabilities of ChatGPT-4 (February 2024 update) in classifying gunshot wounds, specifically distinguishing entrance from exit wounds, and evaluates its potential, limitations, and forensic applicability. Methods: ChatGPT-4 was tested using three datasets: (1) 36 firearm injury images from an external database, (2) 40 images of intact skin from the forensic archive of the University of Catania (negative control), and (3) 40 real-case firearm injury images from the same archive. The AI's performance was assessed before and after machine learning (ML) training, with classification accuracy evaluated through descriptive and inferential statistics. Results: ChatGPT-4 demonstrated a statistically significant improvement in identifying entrance wounds post-ML training, with enhanced descriptive accuracy of morphological features. However, its performance in classifying exit wounds remained limited, reflecting challenges noted in forensic literature. The AI showed high accuracy (95%) in distinguishing intact skin from injuries in the negative control analysis. A lack of standardized datasets and contextual forensic information contributed to misclassification, particularly for exit wounds. Conclusions: While ChatGPT-4 is not yet a substitute for specialized forensic deep learning models, its iterative learning capacity and descriptive improvements suggest potential as a supplementary diagnostic tool in forensic pathology. However, risks such as overconfident misclassifications and AI-generated hallucinations highlight the need for expert oversight and cautious integration in forensic workflows. Future research should prioritize dataset expansion, contextual data integration, and standardized validation protocols to enhance AI reliability in medico-legal diagnostics.