Thirty percent of American teenagers who applied to college used generative AI somewhere in their application essay, according to a 2024 foundry10 survey of more than 500 of them. About half of that group used it to find ideas or to check spelling and grammar; nearly a third had a model write the first draft. The offices reading those essays have meanwhile mostly stopped using the software built to catch them, because false positives and other inaccuracies made AI detectors unreliable. What replaced the detectors is a human reader deciding whether an essay sounds like the person who submitted it.
Policy comes first, and it is not uniform. Brown University forbids applicants to use AI to generate the content of an application, while allowing basic spelling and grammar checking. Caltech's policy for undergraduate and graduate admissions offers a portable test rather than a list: if you are unsure whether what you are about to do is ethical, ask whether it would be ethical to hand the same task to a trusted adult instead of to ChatGPT. A student who has not read the rule at each school is guessing at the one part of the process that has no appeal.
The case against drafting with a model is not only procedural. Researchers at Cornell compared 30,000 human-written admissions essays with text produced by eight large language models. The generated essays were more uniform, more formulaic, and noticeably unlike real student writing. Adding demographic details about a hypothetical applicant to the prompt did not reliably make the output sound more human.
Karen Richardson, dean of admission at Princeton University, writes that the university looks for evidence of curiosity, of intellectual spark, of willingness to step outside the familiar, and of a real desire to learn from other people and about them. The essay's job, in her account, is to show the combination of experience and qualities an applicant would bring to the university, and she has said that an essay written with AI will not be as good or as authentic as one the applicant wrote alone. Robert Penman, executive director of undergraduate admissions at UC Davis, states the same point operationally: an AI-generated answer does not help the office understand the student, their circumstances, their hopes and dreams, their trials and accomplishments. That is the whole function of the document.
The practical tell is internal inconsistency. Readers study grades and test scores alongside the essay. A C in AP Literature attached to prose that reads as though it were prepared for publication gets noticed, and in the worst case it gets read as academic dishonesty. Mechanical sentence rhythm is a separate signal. No software adjudicates any of this: a reader forms a professional judgement and may go back to the high school or to the applicant for confirmation. Once a reader concludes a model wrote the essay, the doubt spreads past the essay to the credibility of the entire application.
Where the tools are genuinely useful is around the essay rather than inside it. Caltech names brainstorming as an ethical use, and a model can generate questions or exercises that start a student looking: what experience shaped you, what you notice that others do not, what story only you can tell. Sam Zhustra, who runs student and family programs at Top Tier Admissions and was previously associate director of admissions at Vanderbilt University, argues that the impulse to reach for AI is usually not an attempt to do less work but confusion and overload in front of the demand to write about yourself. For most applicants the college essay is a genre they never encountered in school, which makes it uncomfortable and unfamiliar.
That also explains where brainstorming goes wrong. Aaron Colton, director of the admissions writing program and assistant professor of English at Pomona College, said in an interview last year that these models are built to reproduce text that at best resembles existing samples of human writing, that they cannot originate ideas, and that they are unlikely to lead students to them. A model also does not know the student, so its questions stay generic unless the student pushes them toward something specific.
The second defensible use is proofreading. Grammarly and Microsoft Editor catch typos, repetition, grammatical errors and clumsy constructions that a student stops seeing on the twentieth read. Flagging a problem and rewriting a sentence are different jobs, and accepting a run of suggested rewrites replaces the student's language with the machine's one sentence at a time. The rule that survives contact with reality: if the tool flags it, fix it yourself.
The third is research, mostly for supplemental prompts that ask why this school. A model can organize intended majors, academic programs, courses, research opportunities and extracurricular interests, and point at what deserves a closer look. It should not be the final source. These tools return outdated information, confuse similarly named programs, and state details confidently that are not true. An essay that praises a professor who has left or a program that no longer exists does not demonstrate interest; it demonstrates that nobody checked. Anything important enough to put in an application is important enough to verify on the college's own site.
Three things are out regardless of policy. Do not ask a model to write, even one paragraph — there is no meaningful ethical distance between a fully generated essay and a few pasted-in sentences, and the pasted sentences make the prose worse anyway. Do not let it adjust facts to improve the story; detail, drama and an arc that did not happen are invention, not craft. And do not disguise ghostwriting as editing: handing over an outline or a rough paragraph and asking the tool to make it better looks like revision, but if what comes back is polished prose the student could not or would not have produced, the tool is the author.
Here is what the guidance does not say about itself. Caltech's trusted-adult test is the most honest sentence in any of these policies, and it concedes the real standard. What admissions offices are defending is not unassisted writing — the same advice that bars a model from drafting a paragraph recommends showing the draft to a teacher, a college counselor, or an experienced admissions professional, and treats structured brainstorming with an adult as the healthy alternative to a chatbot. The line is not between help and no help. It is between help that leaves the student's judgement in charge and help that replaces it, and that distinction is invisible in the finished text. Which is precisely why the enforcement mechanism moved from software to intuition.
The part nobody addresses is what happens to the student the intuition gets wrong. Detectors were abandoned for false positives; the readers who replaced them are running the same pattern-match without producing evidence, and the flagship signal — strong prose, ordinary transcript — is by construction a description of a student whose record understates how well they write. There are three firm rules for what an applicant must not do and none at all for what an applicant does when the reader is mistaken.
The 30% figure is from 2024, and I would not bet on it being the peak. As the share climbs, the readable difference between an essay written with a model and one written with a well-paid adult narrows toward nothing, and admissions offices will be enforcing an authenticity standard whose evidence they have already given up.