Meredith Martin, Professor of English and CDH Faculty Director, and Wouter Haverals, CDH Postdoctoral Research Associate, have published a pre-print revealing a striking pattern: both humans and AI systems show strong bias based on perceived authorship rather than actual content quality. The researchers built a dataset of stylistic rewrites inspired by Raymond Queneau’s “Exercises in Style,” then asked 556 people and a panel of LLMs a simple question: which of these two texts better captures a specific style—the original or the AI’s attempt? Crucially, they manipulated the labels, sometimes telling evaluators the AI-generated text was human-written and vice versa. They found that the same text received significantly different ratings depending solely on its attributed author, with AI models showing 2.5 times stronger bias than humans toward content labeled as human-authored. This suggests AI systems have absorbed cultural assumptions that devalue machine creativity. This research was supported by the Princeton Language and Intelligence (PLI) Seed Grant Program.
Editors’ Choice: New research shows everyone prefers human writers, including AI! – CDH@Princeton
