The Link Between Diverse Words and Original Ideas Is Weakening in the AI-Era College Admissions

Le résumé fourni par la source

Concerns that AI homogenizes human thinking appear at odds with findings that LLM writing is perceived as more creative than human writing. We propose that LLMs enhance surface-level diversity while simultaneously homogenizing underlying ideas. Four large-scale natural experiments tested this preregistered “semantic disjunction” hypothesis in the high-stakes context of college admissions. Comparing personal statements written by 372,793 applicants after-versus-before ChatGPT’s 2022 release, multi-level semantic analyses revealed increasing surface diversity of words simultaneous with decreasing conceptual diversity of sentences (within essays) and whole documents (between essays). Three controlled experiments provided causal evidence that AI use produced semantic disjunction, including in applicants from the natural experiment, and indicated an apparent root mechanism: a positive association between surface and conceptual diversity in humans was weakened or reversed in LLM writing. Despite conceptual homogenization, raters perceived post-ChatGPT essays as more creative. Semantic disjunction explained this apparent paradox: increasing word-level diversity explained increasing creativity judgments, overriding effects of decreased conceptual diversity, even when raters (human and AI) considered many essays together. Avoiding over-reliance on word-level surface diversity by considering multiple semantic levels is likely to support more complete understanding of AI’s effects on human thinking. Disjunctive trends were widespread, but were strongest in minoritized applicants.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.