Can Colleges Tell If You Used AI in Your Essay?

In my other life I'm a software engineer. Some of the best staff-level engineers I know, people I respect immensely, write most or all of their code with AI now. And not just boilerplate. AI can architect systems, debug complex logic, and ship features that would have taken a team days. I know how powerful these tools can be when used by someone who has the skill and experience to get the most out of them. But this admissions cycle, I started noticing something I can't ignore.

This was a strong admissions cycle for our students overall. We saw admits to Caltech, Stanford, UPenn, and other highly selective programs. I say that because I want to be clear: what follows isn't sour grapes. But something has been happening this cycle that caught me off guard, and in conversations with other experienced consultants, I've learned I'm not the only one. Several students with exceptionally strong profiles got rejected early decision at schools where the outcome genuinely shocked me. Not borderline cases at impossible-odds schools. Students with 1580 SATs, rigorous coursework, strong activities, good recommendation strategies, applying to programs where, based on a decade of working with applicants, I would have expected them to be very competitive. Not guaranteed admits. No one is. But well within range.

I want to be careful here. I don't know why any individual student was rejected. Admissions decisions are opaque and multifactored, and even a perfect profile doesn't entitle anyone to admission. But in my work, I get to see the full picture: the transcript, the test scores, the activities, the essays, the recommendation strategy, the school context. I can compare a student's profile to hundreds of others I've worked with over the last ten years. And when several students with that kind of profile get flat rejections in early decision, it's unusual enough that I pay attention to what might have gone differently.

This year, in all of those cases, I know or strongly suspect that the student (against our recommendation) relied heavily on AI during the essay process. I want to be honest: I cannot prove a causal link. Maybe yield protection played a role. Maybe the applicant pool was just that strong. Maybe it was something else entirely. But the pattern gives me pause, and I think students and parents deserve to hear that.

This is one of several reasons I want to have a frank conversation about AI and application essays. Not a lecture. I'm not your parent, and I'm not here to moralize. The admissions system is broken in real ways and has been for a long time. It asks seventeen-year-olds to perform a version of themselves in 650 words, under enormous pressure, while maintaining a GPA, running clubs, preparing for standardized tests, and somehow "demonstrating passion." Wealthy families have always had access to private counselors, essay editors, and test prep that other students can't afford. AI feels like it levels that playing field. It's right there, it's free, and it produces something that looks, on first glance, like a solid essay.

I understand the temptation. But I've seen enough now to believe that this particular shortcut is backfiring on students who might otherwise have been fine. And many of them don't even realize they're doing anything wrong.

The detection question, honestly

You've probably seen the takes online: AI detectors don't work. They produce false positives. They're unreliable. No one can really tell for sure. There's some truth in that. Although there are statistical methods to ivisibly watermark text, AI-generated text generally isn't watermarked. There's no hidden signature that says "a machine wrote this." Detection tools are fundamentally probabilistic: they analyze speech patterns and flag text that statistically resembles AI output. That means they can be wrong.

But most of that discourse is framed around the wrong question. It asks whether AI detectors can prove, with certainty, that a given text was AI-generated. And the answer is: not reliably enough for a court of law.

Fair enough, but that's not the question colleges are asking.

A 2025 working paper out of the University of Chicago Booth School of Business, by researchers Brian Jabarian and Alex Imas, put commercial AI detectors through a rigorous independent evaluation. They built a corpus of 1,992 human-written passages across six genres and generated matched AI versions using four frontier language models (GPT-4.1, Claude Opus 4, Claude Sonnet 4, and Gemini 2.0 Flash). They then tested three commercial detectors (Pangram, GPTZero, and Originality.ai) and one open-source tool (RoBERTa) across text lengths ranging from long passages of roughly 1,000 words down to stubs of under 50 words.

The results were more encouraging for detection than most online commentary would suggest. On medium-to-long passages, the best-performing commercial detector achieved near-perfect separation between human and AI text, with an AUROC (the probability of correctly ranking a random AI passage above a random human passage) of 1.0000 across most genre and model combinations, and never below 0.9979. Its false positive rate was essentially zero at standard thresholds, and its false negative rate stayed between 0.5% and 3.8% depending on the model and threshold used. Even on very short texts under 50 words, where all detectors struggled more, that same tool maintained AUROCs between 0.96 and 1.0. The other two commercial tools formed a second tier: useful on longer texts but weaker on short passages and more vulnerable to evasion.

A 650-word Common App essay falls squarely in the range where the commercial tools performed best.

Now think about what that means in the specific context of college admissions.

A criminal court needs proof beyond a reasonable doubt. A civil court needs a preponderance of evidence. An admissions office at a school with a 5% acceptance rate needs... nothing. They are not adjudicating guilt. They are not obligated to give you the benefit of the doubt. They are reading 40,000 applications and building a class of 1,600. If your essay triggers even a flicker of suspicion, the simplest and most rational thing for a reader to do is move to the next file. They do not need to prove anything. They do not even need to flag it internally. They just need to feel less confident about your essay than they feel about someone else's.

That is the risk calculus students should be running. Not "can they prove I used AI?" but "can they sense something is off, and does that give them a reason to hesitate?" In a process where thousands of strong applicants are turned away, hesitation is enough.

And I know teenagers tend to think adults over 30 are oblivious when it comes to AI. Maybe some of us are. But top universities employ some of the best AI and computer science researchers in the world. Yann LeCun is on the faculty at NYU. Fei-Fei Li is at Stanford. Stuart Russell is at Berkeley. These are people who consult for OpenAI, Anthropic, and the organizations actually building the models you're using. This is not your high school English class. Universities are very aware of what AI can do, and they can afford to build, buy, or license tools and processes that are not available to the general public. They've had several years to work on it. The research I described above is publicly available, based on commercial tools anyone can purchase, and even those perform well enough that I would never advise a student to bet their application on outsmarting them. What a well-resourced admissions office might deploy internally is anyone's guess.

But even if AI detectors don't catch it...

Here's something worth sitting with: the entire point of the application essay is to stand out. That's the job. In a pile of 40,000 applications, your essay exists to make a reader pause, remember you, and advocate for you in committee. AI is architecturally incapable of doing that. It is a next-word-prediction engine trained on enormous volumes of text. Its output regresses toward the mean by design. Asking it to help you stand out in a writing competition is like hiring a cover band to win a songwriting contest. The whole value proposition runs in the wrong direction.

And the person reading your essay is not your classmate. An experienced admissions officer may read a thousand or more essays in a single cycle. Over a career, that number reaches into the tens of thousands. Consultants who have been in this space for a decade have seen hundreds of application files at close range and tracked outcomes across years. That kind of pattern recognition is deep and hard to fool. Students tend to evaluate their own essays in isolation: "does this sound good to me?" But the reader is evaluating your essay against a massive mental database of what authentic student writing sounds like, what AI-assisted writing sounds like, and what the difference feels like on the page. You are bringing your first essay to someone who has read ten thousand of them. That is not an information asymmetry that works in your favor.

What do those readers actually pick up on? Voice consistency, for one. Does the personal statement sound like it was written by the same person who filled out the activities section and wrote the short-answer responses? A sudden leap in sophistication or a shift in register between your 150-word "Why Us" answer and your main essay is a signal, and it's one that requires no software to detect.

They notice the presence or absence of specificity. Real experiences produce concrete details: names, textures, sounds, the specific thing someone said that stuck with you. AI-generated text tends to gesture at categories of experience ("I developed a profound understanding of socioeconomic inequality") without ever landing on a single moment.

And they notice what I'd call lived texture: the details a student would never think to invent. The way the fluorescent light buzzed in the robotics lab. The fact that your debate partner always clicked her pen before rebuttals. The song that was playing when you changed your research topic. AI doesn't know these things. You do.

Think of it like this. A passport photo and a candid snapshot both show your face. But the passport photo, taken under fluorescent lights with a neutral expression, tells a viewer nothing about who you are. The candid, maybe a little blurry, mid-laugh at your grandmother's kitchen table, tells them something real. AI writes passport-photo essays. Admissions readers are looking for the candid.

The students who don't think they used AI

This is the part that worries me most, because a lot of the public conversation about AI and essays focuses on the obvious case: a student who prompts ChatGPT to write their entire essay from scratch. That happens, and it's a problem. But it's not the most common scenario I see.

More often, it looks like this: a student writes a genuine first draft, one with real memories and real voice. Then they need to cut it from 800 words to 650. So they paste it into ChatGPT and ask it to tighten things up. Or they have a paragraph that feels clunky, so they ask AI to "rephrase this more clearly." Or they want a stronger opening line, so they ask for five options and pick one. Or they write their own draft but run it through AI "just to polish the grammar," and accept a dozen small changes to phrasing along the way.

Each of these feels minor. None of them feels like cheating. But the cumulative effect can be devastating to an essay's voice.

Here's what happens. The AI, in the process of "helping," quietly sands off exactly the things that made the writing distinctive. It replaces an awkward but honest sentence with a smooth but generic one. It swaps a specific, strange detail for a more "readable" alternative. It normalizes the rhythm. It inflates the vocabulary just slightly. And the student, comparing the two versions, thinks the AI version sounds "better" because it sounds more polished. But polished and distinctive are not the same thing. The essay that comes back is cleaner, yes. It's also emptier.

The result is an essay that sits in an uncanny valley. It's not fully AI-generated, so it doesn't read like a chatbot wrote it. But it's not fully human anymore either. It's been smoothed into something that a reader can't quite connect with, that feels slightly off in a way that's hard to articulate but easy to sense. I've read hundreds of these at this point. You develop an instinct for it.

And here is the part students don't think about: even those minor uses may violate a school's integrity policy. Many colleges now address AI use in their application instructions, and some of those policies are broad enough to cover editing assistance, not just full drafting. "This essay should be your own work" does not come with an asterisk that says "except for AI-assisted rephrasing." Policies vary, and you should read each school's instructions carefully. But "I only used it a little" is not a defense you want to rely on.

A safer way to use AI, if you're going to

I'm not going to pretend you'll never open ChatGPT or Claude. But there's a clear line between uses that carry lower risk and uses that will get you in trouble.

Lower risk: asking AI to generate brainstorming questions about your topic. Using it to help structure an outline after you've identified your own angle. Running a finished, genuinely self-written draft through a grammar check. Asking it to flag places where your argument gets muddy.

High risk: having AI generate a full draft or substantial portions of your essay. Asking AI to rephrase, tighten, or "polish" your writing (this is where most students get into trouble without realizing it). Asking AI to mimic a writing style. Having AI invent anecdotes or details you didn't live. Using AI to reverse-engineer "what admissions officers want to hear."

The guiding principle: AI should be a mirror, not a ghostwriter. It can help you see your own ideas more clearly. It should not touch the words themselves. And always read each college's specific instructions. If a school says no AI, that means no AI.

The real fix

The admissions system's dysfunction creates a feedback loop. Students try to reverse-engineer what colleges want. They build activities for appearances. They craft narratives around perceived formulas. They manufacture a "spike." And it usually backfires, because admissions readers, whose entire job is evaluating authenticity, can tell.

Genuine interests, even weird or niche ones, come through in ways that are very hard to fake. The student who spent two years cataloging lichen species in a local park. The one who taught herself bookbinding because she liked the feel of the paper. The one who started a neighborhood chess league because he wanted someone to play with. Those applications feel different because they are different.

"Radical authenticity" is the phrase I keep coming back to. It's not a strategy. It's a posture. It means writing from what you actually know, what you actually feel, what you're actually curious about. And in my experience, it tends to outperform overly coached, overly strategic writing at exactly the schools where students feel the most pressure to perform.

Start early. June before your senior year is not too soon to begin paying attention to your own life. Not to start drafting, just to start noticing. What catches your attention? What do you keep coming back to? What would you do even if nobody were watching and it didn't "count" for anything?

That's harder than typing a prompt into a chatbot, but it actually works in our experience.

Ready to transform your educational journey? Sign up for a free intro session with us today to discover our unique teaching philosophy and how we can help you achieve your academic dreams.