Your teen signed a fraud statement before they clicked submit.
Not metaphorically. The Common App makes every applicant electronically sign a statement that everything in the application is their own original work. And since 2023, the Common App's fraud policy has named the thing parents keep asking us about: submitting "the substantive content or output of an artificial intelligence platform, technology, or algorithm" as your own counts as fraud. That one rule covers every one of the more than 1,000 colleges on the Common App, whether or not a given school has written its own AI policy.
So the honest answer to "do colleges check essays for AI" is: most of them don't need to check. You already told them it was yours.
The signature your teen signed
Buried in the final screens before submission, every Common App applicant affirms that the application is their own work, factually true, and honestly presented. Yale's admissions office spells out what that means on its own AI policy page: submitting personal statements composed by text-generating software "may result in admission revocation or expulsion."
Revocation is the word to sit with. This is not a "the essay might be weaker" situation. An admission offer is a conditional promise, and one of the conditions is that the application was honestly presented. If a school later decides it wasn't, the offer can be pulled. Months later. After the deposit.
There is a sensible line in these policies, and it matters. Yale's page says using AI to review grammar or spelling, or to get general advice or topic suggestions at the start of the writing process, does not count as fraud. The line is between help with your writing and replacing your writing. Proofreading your draft is fine. Having the machine write the draft is fraud.
What the schools actually say
Beyond the Common App baseline, a handful of schools have written their own rules, and they do not all say the same thing. That inconsistency is itself the risk: your teen is applying to a dozen schools with a dozen different expectations.
Georgetown takes the hardest public line. Its undergraduate application affirmation says the use of AI tools "to complete any portion of the application, including essays" is prohibited, and false information can lead to rescission or dismissal. Georgetown also just joined the Common App in a three-year pilot, so this cycle's applicants can choose either system. The ban applies either way.
Brown draws the line at content: AI for basic proofreading, spelling, and grammar is fine, but the content of every essay, short answer, and other material must be the applicant's own work.
Caltech asks applicants to read its ethical-use guidelines and, in some cases, disclose how they used AI. Brainstorming and grammar checks are allowed. Outlining the essay, drafting text, or translating it from another language are not.
Cornell allows AI for idea generation but prohibits using it for drafting or editing.
Yale, as noted above, treats AI-composed application content as fraud under the Common App policy, with revocation or expulsion on the table, while explicitly allowing grammar checks and early-stage topic help.
Most schools, though, have said nothing at all. One review of more than 170 universities found a majority still have no published AI policy. That silence does not mean permission. It means the only rule on the books is the Common App's, and it is a fraud rule.
Why detectors are the least of it
Parents ask us whether colleges run essays through AI detectors, and the answer is: some do, but the detectors are not the thing to worry about. A Cornell study this year found detectors are remarkably accurate on a clean, single-pass AI essay, and then fall apart once a student revises and rewrites the draft in their own words. Which is exactly what a careful student would do.
The field is starting to admit this. New guidance for admissions offices urges them to avoid unreliable detection tools, partly because the false-positive problem is real: an actual student's writing can look like AI, especially a strong student who writes cleanly. Nobody should be manufacturing typos to look human. The goal is the opposite: keep the quirks, the specifics, the rhythm, and the observations that make a piece of writing recognizably one person's.
What admissions readers actually rely on is much older technology: the rest of the file. An essay's voice gets checked against recommendation letters, short answers, and interviews, whether anyone runs a detector or not. An AI-polished essay next to a teacher letter that describes a completely different writer is the mismatch that gets noticed. Not a score. A person.
The watermark sleeper
Here is the part most families have not heard. Since August 2026, new Claude models have carried an invisible statistical watermark in generated text. It is not hidden characters and it cannot be seen or stripped by reformatting. It works by nudging the model's word choices in a detectable pattern, which means the mark lives in the wording itself and survives copy-paste.
Today, almost nobody can check for it. Detection sits in private preview, with no public tool a college could use on an essay right now. Full paraphrasing also removes the mark. So this is not a warning that your teen's essay will be watermark-checked tomorrow.
It is a warning about the direction of travel. The marking exists because the EU's AI Act now requires it, and Anthropic applied it worldwide. The text your teen pastes out of a chatbot today can carry a machine-readable signature that did not exist a year ago. The sleeper risk is not that colleges are checking watermarks this cycle. It is that families are still behaving as if AI text is untraceable, and that assumption has an expiration date.
The second risk: AI is reading too
While students have been told to be careful about using AI to write applications, colleges have quietly started using AI to read them. About half of admissions offices now use AI somewhere in their review process, and roughly six in ten of those use it to review personal essays. Virginia Tech pairs every essay with a human reader and an AI scorer; when the two disagree, a second human reads it. UNC Chapel Hill uses AI for initial screening before officers ever see the essays.
This is the second risk vector, and it is the one parents miss. The same machinery that scores essays also summarizes files and flags inconsistencies. An AI-assisted essay is not the only exposure; an activities list with inflated titles or vague claims can get cross-checked the same way. Admissions AI does not just read your teen's essay. It reads whether the whole file agrees with itself.
That is why our application scorecard has an entire dimension for this: Essays x Activities Agreement. Does the person in the essays match the person in the activities list? A file that agrees with itself survives both human readers and machine ones. A file that doesn't is exactly what both are now built to catch.
Three safer moves
1. Write it yourself, and keep the drafts. The safest essay is the one your teen actually wrote, revised multiple times, with a version history to prove it. Drafts are not just good process; they are evidence, if anyone ever asks.
2. Use AI like a tutor, not a ghostwriter. Brainstorming, grammar checks, and early topic help are within the lines at most schools. Generating sentences, paragraphs, or the essay itself crosses it. When in doubt, read the school's own page; the strict ones (Georgetown, Brown) have put it in writing.
3. Check that the whole file sounds like one person. Read the essay next to the activities list and the short answers. If they don't sound like the same teenager, fix that before submitting. Consistency is what both human readers and admissions AI are checking for.
Start with the free scorecard
Not sure whether your teen's essay sounds like them? Send it for a free scorecard. A former Kellogg admissions reader scores the essay and emails all six scores within 1 business day, including whether the voice sounds like your teen, plus the one red flag we would raise first.
Curious what a scored audit actually looks like? See a representative sample audit - the same format every client receives.