Readers Prefer AI Trained on Copyrighted Books over Expert Human Writers (arxiv.org)

A recent study has found that readers increasingly prefer outputs from AI models trained on copyrighted books over texts written by expert human authors. The research involved comparing writings from MFA-trained writers with AI models, such as ChatGPT and Claude, which were prompted to emulate the styles of 50 award-winning authors. Initially, AI-generated texts were rated poorly by MFA readers for stylistic integrity and quality. However, through fine-tuning the AI on complete works of specific authors, the results dramatically shifted—both MFA and general readers rated the outputs significantly higher in fidelity and quality. This fine-tuning also resulted in outputs that were largely undetectable as AI-generated by current detection tools. This study is significant for the AI/ML community as it challenges traditional views on AI-generated content and its place within the literary field. The ability of fine-tuned models to produce human-like writing at a fraction of the cost—99.7% less than traditional author compensation—raises important questions about copyright and fair use, particularly regarding the potential transformation of AI as a content creator. Additionally, the findings suggest that enhancing AI's ability to closely mimic specific writing styles not only improves quality but also circumvents legal concerns regarding originality, providing new avenues for the creation of derivative literary content.