The finding
Researchers analyzing nearly 7,500 applications to a U.S. public-policy graduate program found that a majority of 2025 applicants submitted at least one essay classified as primarily AI-generated, even though the program prohibited that use. arXiv: AI-written admissions essays are widespread but penalized.
A study of nearly 7,500 applications reports better surface-level writing after ChatGPT arrived, but lower admission rates when essays appeared primarily AI-generated. arXiv: AI-written admissions essays are widespread but penalized.
What the evidence actually shows
The paper uses the introduction of ChatGPT as a timing break, compares writing before and after that point, and adds an experiment with admissions officers. The researchers report that AI availability improved writing quality while applications with AI-written essays were admitted less often than comparable applications; officers also rated writing lower when they believed AI had produced it. arXiv: AI-written admissions essays are widespread but penalized. University of Pennsylvania Admissions: Guidance on generative AI and authenticity.
The decision hidden inside the headline
The consequential finding is not that an algorithm can perfectly identify AI. It is that evaluators may react to perceived authorship, flattened voice and rule compliance at the same time. For applicants, the risk therefore exists even when prose is polished: an essay can become technically smoother while offering weaker evidence of judgment, memory and individual perspective. arXiv: AI-written admissions essays are widespread but penalized. University of Pennsylvania Admissions: Guidance on generative AI and authenticity.
What the record cannot yet prove
The dataset comes from one graduate public-policy program, not the Ivy League undergraduate pool, and any classifier can make mistakes. The study cannot establish that every college uses the same standards or that perceived AI authorship alone caused every adverse decision. University of Pennsylvania Admissions: Guidance on generative AI and authenticity.
What applicants should do now
Use AI, if permitted, for bounded brainstorming or proofreading rather than replacing the student’s reasoning and language. Follow each college’s stated policy, keep drafts and notes, and make sure every example can be discussed naturally in an interview. arXiv: AI-written admissions essays are widespread but penalized.
Preserve the exact cycle, program and applicant category in any checklist. When a secondary report and a controlling admissions page differ, the current first-party instruction governs the application. arXiv: AI-written admissions essays are widespread but penalized.
SOURCES
Documents, reporting and comparative context
Source key: Tier 1 is a direct document, official publication or first-party source. Tier 2 is independent reporting or analysis. Tier 3 may identify a lead but cannot establish a factual claim on its own. Comparative-context sources clarify terminology or provide a parallel case; they do not prove the central finding.
UPDATES & CORRECTIONS
Published Sept. 26, 2026: Added after a source review and duplicate check against The AO archive.
