Caltech tells applicants to ask a simple question before using AI: could a trusted adult help you with this task? Yale's admissions office is blunter, stating that breaking its AI policy can lead to an admission decision being revoked or a student being expelled. Between the heuristic and the threat sits a number that makes both of them awkward: 84% of high schoolers had used AI by the middle of 2025, and more than 85% of students use it for schoolwork. The rules governing AI in college applications are written school by school, verified by almost nobody, and already being tested by nearly everyone applying.
The operative rule is not a principle but a lookup. Some colleges and universities prohibit AI outright in any part of an application. Others permit it for brainstorming or for checking prose. Because there is no shared standard, the first real step is reading the applicant section of every school on the list, before writing anything rather than after. The one line that survives the variation is this: AI can support a student's thinking, but it cannot replace it.
The least contested use sits at the very beginning, before an essay exists. Straightforward prompts let Claude or ChatGPT work through the sites of hundreds of colleges and universities to pull test requirements, undergraduate enrollment, location, available majors, and room and board costs into one comparable set. That removes dozens of hours of manual searching and produces a preliminary list worth investigating properly. The failure mode is accuracy. Models return wrong or stale figures, and test requirements, application deadlines, financial aid rules and academic programs all change from year to year. Every number that comes out of a model has to be checked against the school itself.
Brainstorming is where the line gets genuinely thin. Caltech's trusted-adult test is useful here: a teacher or parent would plausibly ask leading questions about the events that shaped you, so asking a model to conduct that interview can be acceptable and productive. But asking questions and knowing which answer deserves attention are different skills. An experienced teacher or counselor notices the unexpected detail, sees the theme running through unrelated episodes, pushes back on an answer that is too smooth, and judges how a topic will sit inside the application as a whole. A model has none of that context.
Which is why there is a real gap between "What questions should I consider while brainstorming?" and "Suggest five interesting college essay topics based on these facts about my life." The first keeps the student's judgement at the center. The second hands over the self-reflection the essay is supposed to demonstrate. A prompt that stays on the right side of that line:
"Interview me about an experience that mattered to me. Ask one question at a time that will help me uncover specific details, contradictions or ideas I haven't considered. Don't suggest an essay topic or write anything for me."
There is a second constraint that gets less attention. Generative models gravitate toward familiar language and familiar narrative shapes. Ask one for unique college essay topics and the suggestions come back resembling each other, because the model is working from patterns distilled across an enormous body of text. A good personal essay runs in the opposite direction. The application is closer to a short memoir, and its value lives in the specific: the odd detail, the unexpected connection, the observation that could only belong to one person.
Editing is the other permitted zone, and the distinction that matters is between finding a problem and producing the fix. "Rewrite this paragraph to make it more sophisticated" makes the model the author. "Identify any grammatical errors, repetitive language or sentences that are difficult to understand. Don't rewrite anything. Explain the problem and let me revise it myself" makes it an editor marking the margin. The difference looks small on the page and is not. Even then, every suggestion needs review, because a model can introduce an error, sand off an individual style, or correct a deliberately colloquial line. A college essay is not a research paper, and flawless smoothness is not what makes one memorable.
The safest applications of all are administrative: turning a list of deadlines into a working plan, organizing notes on schools, building a comparison of financial aid offers. Admissions work splits cleanly into two kinds. The administrative kind is organizing, searching, scheduling and comparing. The evaluative kind is saying what matters to you, describing what you achieved, explaining how you think, and rendering experience in your own voice. AI is genuinely good at the first and should be kept at arm's length from the second.
Here is what stands out to me across all four of those uses: three of them happen outside the application. The only places where AI is unambiguously safe are the places admissions officers never read. That is not really guidance about tools. It is a quiet acknowledgment that the personal essay has become a fragile instrument — an artifact whose entire value depends on an authorship claim that the institution reading it has no reliable way to verify.
Because detection does not work. AI detection tools exist and perform imperfectly, and any site promising a definitive verdict on a text's origin deserves skepticism; research and reporting have repeatedly documented false positives and reliability problems. What experienced readers actually catch are different signals: an abrupt change in voice, language that stays too general, vocabulary that does not match the rest of the file, an essay that looks nothing like the student's earlier work. Research covered by Forbes indicates that experienced people can recognize the characteristics of generated text. No single detection tool is used across all colleges, and schools generally decline to say how they screen at all.
That asymmetry is the part of this nobody frames honestly. A rule that cannot be verified is enforced primarily against the applicant conscientious enough to look it up. Yale's rescission language is real, but it is predicated on finding out, and the finding out rests on a reader's trained eye rather than on any system. The student who never checks the policy faces roughly the same odds as the one who spends an evening reading twelve of them.
Which pushes the whole question back where it started, onto the applicant. If disclosure is requested or an attestation is required, it gets answered fully and honestly; if a school does not ask about permitted use, there is usually no reason to volunteer it in the additional information section, which is better spent on circumstances the rest of the application leaves unexplained. The tests that actually mean something are whether a teacher who knows your writing would recognize the essay as yours, whether you could explain why you chose each example, whether you could discuss the ideas naturally in an interview, and whether the vocabulary and conclusions match how you actually think. Those are human standards, and they are the only ones holding.
Eighty-four percent of high schoolers had used AI by mid-2025. Admissions policies are still written as though that number were small enough to police. The distance between what the rules forbid and what anyone can confirm is now the space the entire process operates in.