The Digital Divide in Generative AI: Evidence from Large Language Model Use in College Admissions Essays

A Cornell University research team led by Jinsook Lee and Rene Kizilcec published groundbreaking evidence of generative AI's complex impact on educational equity. Analyzing 81,663 de-identified applications to a selective U.S. university from 2020-2024, they developed a distribution-based detector trained on both synthetic and historical essays to estimate LLM use. The study reveals a dramatic surge in AI-assisted writing, with estimated adoption doubling in 2024 across all socioeconomic groups.

Technical analysis shows lower-SES applicants (using fee-waiver status as proxy) adopted AI tools at disproportionately higher rates, consistent with an 'access hypothesis' where LLMs substitute for scarce writing support. Surface-level linguistic features converged post-2023, with the most significant changes occurring among fee-waived and rejected applicants. However, the critical finding reveals an equity paradox: while AI adoption increased for all groups, each percentage increase in estimated LLM use correlated with approximately twice the decline in predicted admission probability for lower-SES applicants compared to their higher-SES counterparts, even after controlling for academic credentials and writing style metrics.

This research provides the first large-scale longitudinal evidence linking LLM adoption to actual evaluative outcomes in high-stakes contexts. The findings challenge assumptions about AI as an equalizer, suggesting instead that AI-assisted writing may amplify existing inequalities when used in evaluative contexts. The study raises urgent questions about the validity of essay-based assessment in the AI era and highlights how technological access alone doesn't guarantee equitable outcomes when institutional evaluation systems may penalize certain uses of AI assistance.