AI Replicates Literary Styles with Remarkable Accuracy

A joint study by Stony Brook University and Columbia Law School demonstrates that AI models can convincingly emulate the writing styles of celebrated authors after being trained on merely two books. The research tested 50 distinguished writers including Nobel laureate Han Kang and Booker Prize winner Salman Rushdie.

Methodology and Findings

The team employed two primary training approaches:

1. Context prompts using GPT-4o, Claude3.5Sonnet, and Gemini1.5Pro
2. Author-specific fine-tuning (exclusively via GPT-4o API)

159 participants (28 writing experts and 131 lay readers) evaluated text samples on Prolific without knowing their origin. Key discoveries include:

• Experts initially preferred human-written texts by significant margins
• After fine-tuning, expert preference for AI-generated style increased 800%
• Writing quality selections favoring AI doubled post-fine-tuning
• Modern detection tools identified only 3% of fine-tuned outputs as AI-generated

Implications for Publishing and Law

The study reveals several critical insights:

• Training volume showed no correlation with output quality - two books proved sufficient
• Evaluation convergence between experts and non-experts suggests broad acceptance of AI quality
• Cost analysis shows stark contrast: $25,000 for professional writing vs $81 for AI training

The findings emerge amid ongoing U.S. copyright lawsuits regarding AI's use of protected materials. Researchers propose legal distinctions between general AI output and targeted author imitation.

Key Points:

• ✍️ Style replication: AI achieves convincing author mimicry with minimal training data
• 📈 Quality recognition: Expert preference shifts dramatically toward AI after fine-tuning
• ⚖️ Legal ramifications: Study informs active copyright debates about fair use boundaries
• 💰 Economic impact: Massive cost differentials may disrupt professional writing markets