Researchers from Stony Brook University and Columbia Law School conducted an experiment in which AI models learned the style of 50 famous writers, including laureates Han Kang and Salman Rushdie. Professional authors and three major AI systems — GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro — created texts in the style of these writers, and 159 participants evaluated the excerpts without knowing who wrote them.
For basic comparison, “in-context prompting” was used, and for deeper style imitation, “fine-tuning” of GPT-4o on two books by each author was applied. Readers chose which text was better in style and quality, having a fragment of the original work for comparison. Each excerpt was evaluated by several people to ensure reliable results.
The results showed that after “fine-tuning,” experts preferred AI texts eight times more often for style and twice as often for writing quality. Almost all such excerpts remained undetected by modern AI-text detection tools. The amount of training data had little impact on the AI’s ability to mimic style: authors with two books were copied as successfully as those with dozens of publications.
Researchers estimated that training AI on the style of one author costs about $81, whereas a professional imitator would charge over $25,000 for such work. They noted that basic AI texts often contain clichés and excessive politeness, but targeted training almost eliminates these issues.
