Study Finds Low-Income College Applicants More Likely to Use AI-Generated Essays

The Emergence of AI in College Application Essays

A new wave is sweeping through the college admissions landscape: artificial intelligence tools like ChatGPT and Claude are now commonplace aids for crafting personal statements and supplemental essays. What began as a novelty following the late 2022 launch of advanced large language models has evolved into a standard practice among high school seniors vying for spots at selective universities. These tools promise to level the playing field by offering free or low-cost writing assistance, but recent research paints a more nuanced picture, particularly for applicants from low-income backgrounds.

Across the United States, selective institutions receive tens of thousands of applications annually, each accompanied by essays meant to reveal an applicant's unique voice, experiences, and aspirations. As AI adoption surges, admissions officers grapple with distinguishing genuine narratives from machine-generated prose. This shift raises profound questions about authenticity, equity, and the very purpose of the college essay in holistic review processes.

A Groundbreaking Study on AI and Socioeconomic Disparities

At the forefront of this discussion is a comprehensive study titled "The Digital Divide in Generative AI: Evidence from Large Language Model Use in College Admissions Essays," led by Jinsook Lee, a Ph.D. candidate at Cornell University, alongside co-authors including AJ Alvero from Cornell's sociology department and researchers from Carnegie Mellon University. Published on arXiv in February 2026, the analysis draws from a de-identified dataset of 81,663 essays submitted to a highly selective U.S. university between the 2019–2020 and 2023–2024 admission cycles.

The researchers partitioned the data into pre- and post-ChatGPT eras, using fee-waiver status as a reliable proxy for lower socioeconomic status (SES). This approach allowed them to track linguistic shifts and correlate them with applicant demographics and admission decisions, providing the first large-scale longitudinal evidence of AI's impact on admissions equity.

Methodology: Detecting AI Fingerprints in Essays

Detecting AI-generated content isn't straightforward, as modern tools produce remarkably human-like text. The study employed a sophisticated distribution-based detector inspired by recent advancements in stylometric analysis. They generated synthetic essays using GPT-4o to mimic human distributions, then quantified LLM usage with an alpha-hat (α̂) score ranging from 0 (no AI) to 1 (fully AI-generated). This score compared token-level likelihoods against human and synthetic reference corpora.

Additional metrics included lexical diversity (type-token ratio, MTLD), syntactic complexity (Flesch Reading Ease), and vocabulary richness (Yule’s K). Post-2023, essays showed convergence in these surface-level features, signaling widespread AI influence. The analysis controlled for GPA, test scores, demographics, and school type to isolate AI's effects.

Key Statistics: Low-Income Applicants Lead in AI Adoption

The numbers are striking. In the 2023–2024 cycle, lower-SES applicants exhibited a mean α̂ of 0.102, compared to 0.080 for higher-SES peers—a 28% higher rate. High-intensity usage (α̂ > 0.13) was overrepresented among low-income applicants by 4 percentage points (22.7% vs. 18.7%). Lower-SES AI essays were notably shorter (607 tokens vs. 628), less diverse (lower MTLD by 5.1%), and more repetitive (higher Yule’s K by 3.3%).

Pre-GPT admission rates were 12.9% for lower SES vs. 23.6% for higher SES. Post-GPT, the gap widened to 14.0 percentage points, with lower-SES rates dipping while higher-SES rose.

Why Low-Income Students Are Turning to Free AI Tools

For many low-income high schoolers, AI represents a rare equalizer. Unlike affluent peers who access private counselors, essay coaches, or premium writing services costing hundreds per hour, lower-SES students often lack such supports. Public schools overburdened by large caseloads provide minimal individualized feedback, leaving students to rely on freely available tools like basic ChatGPT.

Lead author Jinsook Lee noted, “Lower-income students might only be able to use the free tier... and the quality of the outcome of what free-tier ChatGPT gives us is really poor.” This disparity in tool quality—free versions produce more formulaic, less nuanced text—compounds the challenge. First-generation applicants, overrepresented in low-income groups, face additional hurdles navigating essay expectations without familial guidance.

Admission Penalties: AI Hits Low-Income Applicants Harder

AI use correlates with lower admission odds across the board, but the penalty is steeper for lower SES. Logistic regression showed odds ratios of 0.17 for low-income AI users vs. 0.38 for high-income—a nearly twice-as-severe impact. Even after adjusting for credentials and stylometrics, the SES × AI interaction remained significant (p=0.023).

Mediation analysis revealed essay length and word count explain about 20-25% of the differential penalty, with other features acting as suppressors. Admissions officers may subconsciously penalize homogenized language, mistaking it for inauthenticity more readily in underrepresented applicants.

Universities Deploy AI to Combat Essay Fabrication

Selective colleges are countering with their own AI. Virginia Tech's essay scorer, trained on past rubrics, cross-checks human evaluations on a 12-point scale. Georgia Tech uses AI for transcript verification and aid eligibility. Caltech employs video AI interviews to probe research claims. Tools flag inconsistencies in grammar, style, and repetition, though false positives risk harming non-native English speakers or neurodiverse applicants—disproportionately low-income.

Evolving Policies: What Colleges Allow and Forbid

Most universities now require applicants to affirm non-use of AI for primary writing, per NACAC guidelines emphasizing integrity. Yale permits AI for grammar checks but not content generation. UNC and Virginia Tech blend AI screening with human oversight. A 2025 Kaplan survey found more colleges clarifying rules, though many leave applicants guessing. Ethical use—brainstorming, outlining, proofreading—is increasingly tolerated, but full drafts trigger rejection risks.

Equity Concerns: Widening the Admissions Divide

AI promised democratization but delivers unequal returns. A separate Cornell analysis of 150,000+ essays found AI output mimics privileged, male voices—longer words, less variety—further disadvantaging authentic low-income stories of resilience. This "digital divide" shifts from access barriers to outcome disparities, potentially undermining diversity goals post-affirmative action.

Stakeholders warn of eroded trust: essays lose value as signals of voice if homogenized. Low-income admits, already scarce, could plummet without intervention.Inside Higher Ed coverage highlights calls for prompt redesigns emphasizing multimedia or interviews.

Solutions: Bridging the Gap for Underrepresented Applicants

• Expand free, high-quality writing programs like College Track's AI Essay Lab, which guides ethical use for first-gen/low-income students.
• Institutions offer subsidized premium AI access or human-AI hybrid coaching.
• Admissions pivot to video essays, portfolios, or interviews to capture voice beyond text.
• Transparent AI detection with appeals for flagged low-SES apps.
• Public schools partner with nonprofits for essay workshops, reducing AI overreliance.

Proactive measures, like Gates Foundation-backed chatbots for aid navigation, show promise in equitable AI deployment.

The Road Ahead for Authentic Admissions

As 2026–2027 cycles approach, the essay's role hangs in balance. Experts advocate de-emphasizing text for multifaceted evaluations prioritizing lived experiences. For low-income aspirants, success lies in blending AI ethically with personal iteration—prompting tools with specifics, then rewriting extensively. Ultimately, fostering genuine storytelling ensures colleges welcome diverse talents, not algorithmic echoes.

Photo by Google DeepMind on Unsplash