A recent study from Stony Brook University and Columbia University Law School shows that **artificial intelligence models fine-tuned with small amounts of data can generate works in the style of famous writers that readers prefer, even surpassing human professional imitators.** This finding not only demonstrates the remarkable ability of generative AI in literary imitation but also brings key implications for ongoing copyright lawsuits and debates over "fair use" in the United States.
Research Method: AI Competes with Human Experts
Researchers used three AI systems—GPT-4o, Claude3.5Sonnet, and Gemini1.5Pro—as well as professional writers to create articles in the styles of 50 renowned authors, including Nobel Prize winner Han Kang and Booker Prize winner Salman Rushdie.
In the experiment, the research team employed two AI generation methods:
Contextual Prompting: Using GPT-4o, Claude3.5Sonnet, and Gemini1.5Pro, they provided the same instructions and example texts.
Targeted Fine-Tuning: Only using GPT-4o, which supports API functions, they purchased e-books from 30 authors and conducted targeted training on the model. To the researchers' surprise, even authors like Tony Tulathimutte, who have published only two books, were imitated just as effectively as prolific writers like Haruki Murakami.
Key Findings: Fine-Tuned AI Works Are Preferred by Readers
159 participants, including 28 writing experts and 131 non-experts, evaluated and compared the works side by side without knowing whether the author was human or AI. The results showed:
Basic AI Output: Under basic prompts, expert readers strongly favored human-written texts.
Finetuned AI Output: The results underwent a dramatic change. After finetuning, the probability of experts choosing AI-written articles increased eightfold in terms of style and doubled in writing quality. Both expert and non-expert readers preferred AI-generated texts in terms of style and quality.
AI Detection Fails: Modern AI detectors marked 97% of standard AI outputs as machine-generated, but only 3% of finetuned AI works were identified, demonstrating its "indistinguishable from human" capability.
The study noted that common issues with general AI outputs, such as clichés and unnatural politeness, were significantly improved through targeted training.
Enormous Cost Difference: AI Imitation Costs Just 0.3% of Human Costs
The study also revealed a significant cost difference: training an AI model to adapt to the style of one author costs about $81 per author; while human professionals charge up to $25,000 for the same amount of imitative text, reducing costs by 99.7% with AI.
Far-Reaching Impacts: Literary Imitations Challenge 'Fair Use'
The researchers emphasized that these findings provide crucial arguments for lawsuits currently being heard in U.S. courts regarding how AI companies use copyrighted materials. If readers prefer AI-created imitations, this may constitute a clear "harm" to the original work market, directly challenging the core legal principle of "fair use."
The U.S. Copyright Office has previously warned that even if AI does not copy verbatim, it could still compete with original works. The study authors suggest that the law should differentiate between general AI models and those specifically designed to imitate particular authors, and propose that targeted imitation has little legal basis and should be considered for prohibition, or require clear labeling of AI-generated texts.
