Authorship Attribution of Arabic Criminal Texts Using Large Language Models: A Comparative Evaluation of ChatGPT, DeepSeek, and Gemini

This study investigates the ability of three large language models (LLMs)—ChatGPT, DeepSeek, and Gemini—to attribute authorship of Arabic criminal texts in a zero-shot setting, with no task-specific training or fine-tuning. Using a quantitative experimental design, each model attributes 24 anonymous criminal texts against reference writings from 12 Arabic authors. The results reveal limited effectiveness, with only ChatGPT achieving a statistically significant accuracy rate of 25%, above the 8.3% chance level. These findings demonstrate that current LLMs in zero-shot settings lack sufficient reliability for definitive authorship attribution (AA) of short Arabic criminal texts, highlighting a gap between their general linguistic capabilities and the specific requirements of forensic textual analysis. While LLMs show preliminary potential, their current implementation cannot replace human expertise in high-stakes forensic contexts involving Arabic texts.