A recent research paper published on arXiv cs.CL investigates how effectively six different large language models can mimic the writing styles of specific authors. The models evaluated in the study are Gemini 3 Pro, Claude Opus, GPT 5.2, GPT 5.2 Pro, DeepSeek 3.2, and Grok 4.1 Fast. To test stylistic replication, the researchers utilized short product review texts known for high within-author consistency. The evaluation compared the outputs against training data metrics using cosine distance, employing both simple prompts with minimal instructions and complex prompts containing explicit stylometric guidance.
Impersonation Testing in Modern AI
As large language models become more advanced, evaluating their ability to adopt specific writing personas is increasingly relevant for security, content creation, and stylometric research. By testing models like DeepSeek 3.2 and various iterations of GPT and Gemini with both basic and detailed prompts, the study offers insight into how prompt engineering influences stylistic mimicry. The use of product reviews provides a controlled environment where individual author style remains stable.
Open Questions and Future Metrics
Despite the comparative evaluation of these six models, the announcement did not specify the exact outcomes of the cosine distance measurements or which model performed the most accurate impersonation. It remains unclear how these findings will scale to longer texts or different genres beyond short product reviews. Observers will need to watch for future updates or the full data release from the arXiv cs.CL researchers to understand the practical implications for authorship verification.
Further details regarding the methodology and comparative performance rankings are expected to emerge as the research is reviewed by the wider academic community.
Lead image: cottonbro studio / Pexels
