LLM-Era College Admissions Essays Exhibit Paradoxical Semantic Trends

Concerns that AI will homogenize human thinking—and homogenization evidence when human-AI collaboration is observed—appear at odds with findings that AI-assisted content is judged as highly original. Using a multi-method approach that combined a controlled experiment with two large-scale natural experiments in over 160,000 U.S. college admissions essays, we observed a counterintuitive pattern: AI-assisted essays and essays written after the release of ChatGPT showed greater lexical diversity, but the increasingly diverse words actually expressed increasingly homogeneous ideas both within and across essays. This “paradoxical homogenization” pattern explained higher creativity ratings given to AI-generated essays than more idea-diverse human essays. Paradoxical homogenization was robust across demographic subgroups but emerged disproportionately among racial and linguistic minority applicants. Event-time analyses and robustness checks ruled out pre-existing trends and pandemic-related confounds. Results suggest that surface-level diversity of AI-generated content may mask deeper constraints on the breadth of human ideation. Paradoxical homogenization appears to have reached a real-world, high-stakes educational context where writing is expressly personal and AI use is explicitly prohibited.