Unveiling Identity Deception in Cybercrime: ChatGPT's Mimicry of Human Writing Styles.

This article investigates the capability of ChatGPT, a large language model (LLM), to learn and replicate the unique writing styles of individual authors using a one-shot training method on texts from 50 Amazon product reviewers. The study compares human-written and ChatGPT-generated texts (using versions 3.5 and 4 with simple and complex prompts) through classification accuracy, stylometric distribution, and word-expression associations. Results indicate that ChatGPT struggles to accurately mimic individual writing styles, often defaulting to a limited set of preferred words and expressions distinct from human usage, though ChatGPT-4 shows modest improvement over ChatGPT-3.5. The findings highlight challenges in style imitation with limited training data and suggest that while impersonation of public figures may be more feasible due to available data, broader misuse remains a concern as AI technologies advance. The study underscores the need for ongoing research into detection methods, ethical considerations, and policy responses related to AI-generated text impersonation.