A new study shows that AI models can generate the writing style of famous authors after being fine-tuned with just two books, and readers' preferences for these works even exceed those of professional imitators. The study was conducted by a research team from Stony Brook University and Columbia Law School, involving the creation of styles of 50 renowned authors, including Nobel Prize winner Han Kang and Booker Prize winner Salman Rushdie.
In the study, 159 participants, including 28 writing experts and 131 non-experts, evaluated different text passages through the crowdsourcing platform Prolific. During the evaluation, participants were not aware whether the texts were written by humans or AI. The study used two main AI training methods: context prompts and fine-tuning for specific authors. In the context prompt method, the research team used three major AI systems: GPT-4o, Claude3.5Sonnet, and Gemini1.5Pro, generating texts with the same instructions and sample texts. In the specific author fine-tuning, only GPT-4o supported the required API functionality, so the research team purchased digital copies of books for 30 authors and trained independent models for each author.
Participants were asked to compare two passages of text and choose the one they considered better. For style evaluation, participants also saw an excerpt from the original author. The study results showed that when using basic context prompts, experts significantly preferred human texts, while non-experts' choices were more scattered. However, after fine-tuning, experts' preference for AI-generated texts in terms of style increased eightfold; the proportion of AI texts chosen for writing quality also doubled. Modern AI detection tools can identify standard AI output with up to 97% accuracy, but only 3% for fine-tuned outputs.
Additionally, the study found that the amount of training data did not affect the results. Some authors who had published only two books could still have their styles well simulated. The convergence of evaluation standards between experts and non-experts indicates that the quality of fine-tuned AI texts has been widely recognized. Since professional writing costs up to $25,000, while training AI costs only about $81, the economic benefits are significant.
These study results come at a time when US courts are reviewing lawsuits about how AI acquires and uses copyrighted materials. The research team suggests that laws should clearly distinguish between AI specifically imitating certain authors, possibly requiring a ban on AI copying individual authors' styles or mandating clear identification of AI-generated texts.
Key Points:
📚 AI models can generate the writing style of famous authors with just two books.
🧑🎓 After fine-tuning, readers' preferences for AI-generated texts have significantly increased.
⚖️ The study results may impact U.S. copyright law and debates on fair use.
