AI-written admissions essays are widespread but penalized

AI is rapidly transforming higher education, including the application process, yet relatively little is known about its use and consequences. To help close this gap, we analyze nearly 7,500 applications submitted between 2020 and 2025 to a large public policy master’s program in the United States. We find that in the 2025 admissions cycle, the majority of applicants submitted at least one essay that was primarily AI-generated—despite an explicit prohibition against using AI. Leveraging the abrupt introduction of ChatGPT in November 2022, we find that the availability of AI assistants improved the writing quality of submitted essays. These improvements, however, came with an apparent AI penalty: Applicants submitting AI-written essays were admitted less often than comparable non-users. To help explain this penalty, we conduct an experiment with admissions officers, finding that they can often recognize AI writing and rate essays they believe to be AI-generated lower than essays they believe to be human generated. These findings indicate that AI is changing both how applicants write and how that writing is evaluated, raising questions about whether admissions practices and policies designed for a pre-AI era remain appropriate.

Introduction. When ChatGPT was released in November 2022 [1], it fundamentally altered the way people write. AI-generated writing has since proliferated across a variety of domains, ranging from academic journal submissions, to job applications, to UN press briefings [2–6]. University admissions essays have been no exception [7–9]. Historically, admissions essays have been a channel for applicants to demonstrate their writing ability, motivation, and program fit, beyond grades and test scores [10–12]. Now, however, university admissions offices must contend with the possibility that applicants’ submitted essays do not reflect their abilities or authentic voice [7]. Recent survey evidence suggests that AI use may already be widespread, with 30% of college applicants reporting using generative AI to help write their personal essays, 20% of whom reported using AI to produce a final draft [13]. Similarly, Lee et al. [9] estimated that approximately 8–10% of the text in Common Application essays submitted to a competitive engineering school exhibited linguistic patterns characteristic of LLM-generated content in the year after ChatGPT’s release. Despite the speed, scale, and significance of this change, few universities provide formal guidelines on the use of AI in applications [14], perhaps reflecting a lack of evidence on the prevalence and consequences of AI usage in admissions. Recent work has considered the detectability of AI in admissions materials [15–18], the stylistic properties of AI-generated admissions essays [8, 9, 19], and exploratory analysis of the impacts of AI use on admissions outcomes [9]. But it remains unclear how often applicants are using AI, how AI impacts the quality of their applications, and whether AI use affects applicants’ admission decisions. Without answers to these questions, admissions offices may struggle to set and enforce sensible AI policies. This lack of clarity leaves applicants to make individual choices on the appropriate use of AI, potentially exacerbating disparities in the admissions process. Here we begin to close this information gap by analyzing a novel dataset of nearly 7,500 applications to a large American public policy master’s program submitted between 2020 and 2025—covering three admissions cycles before and three after the release of ChatGPT. Using commercial AI detection tools, we find that AI usage is widespread and growing, with 56% of applicants submitting at least one essay that was likely primarily AI-generated in the most recent application cycle. Comparing pre- and post-ChatGPT essays, we find that the broad availability of AI substantially improved the quality of applicants’ essays, particularly for international applicants. Despite these improvements in essay quality, applicants who submitted AI-written essays were nevertheless admitted at lower rates than otherwise similar applicants who did not use AI. To understand the source of this apparent penalty, we conduct an experiment to measure admissions officers’ ability to detect AI-written essays. We find that not only can admissions officers often discriminate between AI and human-written text, but that they also rate essays they suspect to be AI-written more harshly. Together, our findings suggest that AI use is both common and consequential, at least among the applicants we study, but likely more broadly. These results underscore the need for universities to reevaluate the admissions process in response to this new reality, and to develop sensible and enforceable AI policies for it.

Discussion. Drawing on nearly 7,500 applications across six admissions cycles, we document a dramatic shift in applicant behavior. By 2025—just three years after the introduction of ChatGPT—a majority of applicants submitted at least one essay written primarily by AI, despite explicit prohibitions on its use. Among international applicants, the rate approached 70%. We further find that AI use improved the overall quality of submitted essays, particularly the more mechanical aspects of writing. Yet admissions staff could often recognize AI-written essays and appear to have penalized them. Our findings expose a central tension for university admissions. AI can improve the quality of submitted essays while weakening the connection between those essays and applicants’ unaided writing ability—and potentially subjecting applicants to idiosyncratic penalties for using AI. Even before AI, application essays did not necessarily reflect an applicant’s unaided effort, as students frequently turned to mentors, teachers, and others for help. But AI assistants have likely widened the gap between an applicant’s individual writing ability and the quality of the final submitted product. At the same time, by reducing the burden of writing, AI may help applicants communicate their ideas more clearly, even as uncritical use of these tools may also distort or displace those ideas. Ultimately, universities must clarify what they seek to learn from application essays and reconsider whether the purposes those essays have historically served remain appropriate in a world where AI-assisted writing is becoming standard professional practice. These results and their implications should be considered in light of several limitations. First,

Limitations. For example, in settings where AI use is not expressly prohibited, its prevalence may be even higher and its consequences less severe. Second, our estimates of AI use rely on commercial detectors calibrated to limit false positives, and are thus likely conservative. Moreover, to facilitate interpretation, we consider as “AI generated” only essays that were likely primarily written by AI, excluding less extensive uses of AI that were also prohibited. Third, our automated measure of essay quality is an imperfect proxy for expert judgment. Although it generally tracks ratings by admissions staff, it may systematically miss dimensions of quality that matter in admissions. Finally, our estimate of the apparent penalty associated with AI use may be affected by unmeasured confounding. Our double machine learning approach adjusts for hundreds of covariates derived from submitted application materials, but we cannot rule out omitted factors associated with both AI use and admissions decisions. Looking ahead, our findings highlight the need for institutions to move toward deliberate, transparent policies around AI use.