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

The advancement of generative artificial intelligence has made it increasingly convenient for cybercriminals to craft persuasive texts by imitating the writing style of trusted individuals. With the accessibility of low-cost, user-friendly natural language generation systems such as ChatGPT and the abundance of publicly available personal text data, the mass production of deceptive messages with identity impersonation has become notably simplified, consequently broadening the scope of cybercrime. This study investigates ChatGPT's ability to learn the writing style of individuals and replicate texts by mimicking their style. Employing a one-shot training method, texts authored by 50 individuals are used to train ChatGPT for generating new texts that mimic each individual's writing style. The human-written and machine-generated texts are compared with respect to three questions: (a) To what extent can human-written and machine-generated texts be classified? (b) How are the human-written and machine-generated texts distributed and differentiated within the stylometric space? (c) Are any words and expressions significantly associated with either humans or machines? Two different versions of ChatGPT (Versions 3.5 and 4), along with two different prompts (simple and complex), are used for text generation. The results suggest that mimicking writing styles presents a considerable challenge for the current model of ChatGPT. The model exhibited a tendency to select similar preferred words and expressions across different authors, ChatGPT versions, and prompts. Potential reasons for ChatGPT's limited performance in this regard are discussed, along with possible approaches for enhancing its performance.