AI Wins Imitation Game: Readers Prefer Fanfic Written by ChatGPT (www.theregister.com)

Researchers behind a preprint titled "Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers" tested whether AI can convincingly emulate the styles of famous authors. They recruited 28 MFA writers to produce 450-word imitations of 50 award-winning authors (150 human excerpts) and generated 150 AI excerpts using ChatGPT. In blind pairwise tests with 159 evaluators (28 MFA-candidate experts + 131 lay readers), AI produced by in‑context prompting was initially disfavored by experts. However, after fine-tuning ChatGPT on each author’s complete works, preferences flipped: experts and lay readers favored the fine‑tuned AI for stylistic fidelity and writing quality. The authors note fine-tuning appears to remove detectable AI stylistic quirks (e.g., cliché density), and estimate the median cost to fine-tune and generate a 100,000‑word novel at roughly $81—about a 99.7% cost reduction versus hiring a professional writer. Those technical findings carry immediate legal and market implications: if fine‑tuned LLMs can reliably mimic authors at tiny marginal cost, they could displace human writers and alter the "market effect" factor in U.S. fair‑use analyses. With more than 50 copyright suits already targeting AI training (including high‑profile cases like Bartz v. Anthropic and Kadrey v. Meta), the paper argues courts should revisit assumptions that training on copyrighted texts is fair use—even if outputs aren’t verbatim copies—because the underlying copying may substitute for original works. The study strengthens the argument that fine‑tuning on copyrighted corpora poses both commercial risk and potential liability for model developers.