ChatGPT’s ability to imitate writing styles: an analysis guided by forensic text comparison

This study examines ChatGPT’s ability to learn and replicate individuals’ unique writing styles. Using a one-shot training method, short texts from 300 authors of product reviews were used to train ChatGPT to generate new texts mimicking each author’s style. Likelihood ratio-based forensic text comparison experiments were conducted on same-author and different-author text pairs, involving human-written, machine-generated, and mixed texts. The findings indicate that replicating writing styles remains a significant challenge for ChatGPT. Linguistic analysis revealed that while both humans and machines exhibited preferred words and expressions, these preferences were more strongly associated with the machine.