Does AI essay use hurt admissions chances despite quality gains?

Across nearly 7,500 applications to a large American public policy master's program, submitted between 2020 and 2025, the authors report that AI use is both widespread and penalized. In the 2025 cycle, "the majority of applicants submitted at least one essay that was primarily AI-generated—despite an explicit prohibition against using AI." These figures come from commercial detectors. The introduction gives 56% of applicants submitting at least one essay "likely primarily AI-generated" in the most recent cycle, and the discussion says the rate for international applicants "approached 70%." The penalty is the second half of the claim: applicants who submitted AI-written essays "were admitted at lower rates than otherwise similar applicants who did not use AI."

The design uses the timing of ChatGPT's release as its source of variation, comparing essays from before and after November 2022. The authors find that AI availability "improved the writing quality of submitted essays," especially the more mechanical aspects and especially for international applicants. The penalty is estimated with a double machine learning approach that "adjusts for hundreds of covariates derived from submitted application materials." The discussion presents the two results as one "central tension": AI can raise essay quality "while weakening the connection between those essays and applicants' unaided writing ability." Quality rises, and the same essays are discounted.

Against the nearest notes, this paper gives the policy gap a behavioral test. The university AI policies note argues that public guidance states permission categories more clearly than the evidence standards that would protect what a credential certifies. Here the prohibition was explicit, and a majority of applicants still submitted a primarily AI-written essay. The excerpt does not say why applicants ignored it. The conclusion that institutions need "deliberate, transparent policies around AI use" approaches the same gap from the institution's side. The rhetorical-sensitivity note shows an LLM reviewer's scores moving with framing while reported content is kept fixed. This paper's penalty is tied to suspected authorship, and the excerpt contains no comparison that holds content fixed. The temporal signature note reads AI use from process data rather than output, which is the route this paper does not take.

The excerpt does not establish several things. It reports no false-positive or false-negative rates for the detectors. "Calibrated to limit false positives" describes a design choice, and the authors call their prevalence estimates "likely conservative." The penalty is an association that survived covariate adjustment, and the authors say plainly that "we cannot rule out omitted factors associated with both AI use and admissions decisions." The data come from one program at one institution. The claim that prevalence "may be even higher" where AI is not prohibited is the authors' conjecture, not a finding. The implication is that the penalty is a serious reason to revisit what admissions essays are for and what a policy should enforce. The excerpt does not show that the penalty is unfair, or that AI-written essays signal lower ability.