AI Is Writing the Essays. Colleges Can't Agree on What to Do About It.

About half of college applicants now use AI to brainstorm their admissions essays. One in five use it to generate a first draft. And a study of 81,663 applications submitted to a selective university between 2020 and 2024, conducted by researchers at Cornell, Carnegie Mellon, and USC, found that estimated AI use spiked sharply after ChatGPT's release.

The essays are getting more similar. The study found that after 2023, applicant writing converged in surface-level linguistic features. Sentences became smoother, vocabulary more polished, and voice less distinct. AJ Alvero, a professor in Cornell's sociology department and co-author of the paper, told Inside Higher Ed that the trend is concerning: the essay is supposed to let applicants highlight "the idiosyncrasies of their life, how they became who they are."

AI is erasing exactly the thing the essay was designed to capture. And the schools that read those essays have responded in completely different directions.

Three Camps, No Consensus

Brown University bans all AI use in applications. The policy is explicit: artificial intelligence is "not permitted under any circumstances," except for basic spelling and grammar checks. Violations can result in a rescinded admission offer. Georgetown takes the same position.

The University of California system goes further. UC schools run plagiarism detection and warn that applicants caught submitting AI-generated content can be disqualified from all nine campuses. Not just the one they applied to. All of them.

Then there's the opposite end. Johns Hopkins disabled its AI detection tools entirely after false-positive rates hit roughly 4%. UCLA and UC San Diego made the same call. Across five countries, more than 60 universities have now turned off or refused to enable AI writing detectors, most citing the same problem: the tools aren't reliable enough to use on high-stakes decisions.

The middle ground is where most schools have landed, and it's muddy. Caltech published detailed ethical-use guidelines allowing applicants to use AI for research, brainstorming, and grammar, but not for outlining or drafting. Cornell permits "ethical uses." Carnegie Mellon allows grammar and spelling assistance. WashU lets students check clarity. In each case, the line between acceptable and unacceptable use depends on judgments that are nearly impossible to enforce from the outside.

The Common App updated its fraud policy in 2024 to classify the submission of AI-generated content as fraud. But "AI-generated" is doing a lot of work in that sentence. A student who pastes a ChatGPT draft verbatim has committed fraud. A student who uses ChatGPT to brainstorm five possible essay angles, picks one, and writes the essay herself has not. The distance between those two acts is obvious in principle and invisible in the finished product.

About 70% of colleges have no explicit AI policy at all. Twenty-three percent permit limited use. Seven percent ban it outright.

The Detector Problem

The tools that were supposed to solve this don't work well enough. The best commercially available AI detectors top out at roughly 76% accuracy. That means nearly one in four flagged essays could be a false positive.

The false-positive problem is not evenly distributed. Non-native English speakers and students with highly polished writing are disproportionately flagged. For international applicants or first-generation students who worked with writing tutors, a false flag could mean a rejected application or, worse, a fraud accusation. That risk is why Johns Hopkins, UCLA, and dozens of other schools stopped relying on automated detection.

Most colleges that still use detectors treat a flag as a trigger for human review, not an automatic rejection. But human review introduces its own subjectivity. An admissions reader looking at a flagged essay is primed to see AI in the prose. The flag changes how the essay is read even if the student wrote every word.

Who Gets Hurt

The Cornell/Carnegie Mellon study surfaced a finding that complicates the easy narrative about AI as a tool of the privileged. The biggest increases in estimated AI use came from fee-waived applicants, a proxy for lower socioeconomic status. These students are more likely to lack access to private college counselors, essay coaches, and the informal networks that help wealthier applicants polish their writing. AI fills that gap.

But it fills it badly. A separate Cornell study comparing 30,000 human-written essays to AI-generated ones found that large language models produce "highly uniform text" regardless of how the prompt is adjusted for a student's background. Even when researchers specified a person's race, gender, and geography, the output was generic and easily distinguishable from real student writing.

The students using AI the most are the ones who can least afford to sound generic. At a school reading 50,000 applications, a polished but indistinct essay disappears into the pile. The voice that would have made a low-income applicant stand out is the exact thing AI replaces.

And none of this helped with outcomes. The study found that AI use was not associated with improved admissions results. The students who used AI were rejected at the same or higher rates as those who didn't.

What a 17-Year-Old Is Supposed to Do

The practical problem for applicants is that the rules are different at every school, the enforcement is inconsistent, and the consequences range from nothing to disqualification across an entire university system. A student applying to ten colleges in the 2026-27 cycle could face three or four different AI policies, each with different definitions of acceptable use.

Brown says no AI at all. Caltech says AI for brainstorming is fine but not for drafting. Johns Hopkins doesn't check. The UC system checks and will throw you out if they catch you. The Common App says AI content is fraud, but the form doesn't ask you to specify how you used it.

No one has agreed on what the rules are. The only consistent advice that holds across every policy is the simplest: write it yourself.