Can AI fool a forensic linguist? Detecting AI impersonation of an individual's language

With the rapid advancement of large language models (LLMs), concerns have grown regarding their potential misuse, particularly in impersonating individuals through mimicked writing styles. This study explores whether state-of-the-art LLMs, when guided by various prompting techniques, can convincingly imitate an individual’s linguistic identity and potentially deceive forensic authorship verification methods. Using GPT-4o as the model and the Enron email dataset as the corpus, we evaluated four prompting strategies: naïve direct prompting, system-user prompting, self-prompting, and Tree-of-Thoughts (ToT) prompting (Chen and Moscholios, 2024). We assessed generated outputs using multiple authorship verification tools, including n-gram tracing, the Impostors method, LambdaG (Nini et al., 2024), and AdHominem (Boenninghoff et al., 2019). Our findings indicate that while LLMs can approximate an individual’s writing style on a superficial level, forensic linguistic methods remain effective, especially when content-related lexical cues are masked. The results underscore the resilience of authorship verification techniques and suggest that authentic linguistic individuality remains difficult to convincingly replicate by prompting an advanced language model.