Blog · Aug 2026 · 7 min read

The AI Rules Are Out. Read Them Literally.

For the first time, most top business schools will tell you in writing what AI may and may not do to your essays. The mechanisms differ — bans, checkboxes, footnotes, detection software — but the line they draw is the same one. The applicants who get in trouble this cycle will be the ones who didn't read it.

Two years ago the question "can I use ChatGPT on my MBA essays" had no official answer. This cycle it has a dozen, and they are worth reading in their exact words, because the wording is the policy.

Stanford GSB has the hardest line in the field: its application terms call it "improper and a violation of the terms of this application process to have another person or tool write your essays," with denial or revocation of admission as the stated consequence. Note the construction: AI is regulated the same way a ghostwriter is. Harvard asks a yes/no question about whether you used AI; answering yes triggers a short statement, around 75 words, describing how. Kellogg permits AI as a tool and requires a citation at the end of the essay naming it. London Business School wants the same disclosure as a footnote, excluded from the word count. Michigan Ross goes furthest into formality: AI assistance gets an APA-style in-text citation, as in "(OpenAI, personal communication, September 1, 2024)."

Then there is the enforcement tier. Wharton states it may deploy "proprietary and licensed AI detection tools," with flagged applications subject to investigation. Duke Fuqua's application instructions say essays "may also be reviewed using AI-detection tools," and draw the use line in one clean sentence: "AI should help refine your content, not generate it." Columbia writes the line into its Honor Code: generative AI is permitted "for idea generation and/or to edit a candidate's work," but "using these tools to generate complete responses violates the Honor Code," alongside standing language that offers are rescinded for misrepresentation.

Step back from the mechanisms and the convergence is striking. One tracker counts 45 of 117 business schools with explicit AI policies as of this spring, and across bans, disclosures, citations and detectors, they all draw the same boundary: AI may assist your thinking and tighten your prose. It may not be the author. The differences are about how each school intends to catch the difference.

What the officers say when you ask them directly

This August, Clear Admit put the AI question to admissions officers at more than twenty schools, and the answers are more useful than the policies.

From Booth's admissions office: "The more that you put into them, you're going to sound less like yourself… We're not admitting ChatGPT to the program." From Cornell Johnson's, on what AI-drafted essays have in common: "They all look exactly the same." Yale SOM's admissions office offered the most practical rule in the set: "Use it like you would use a trusted friend. So you would never have a friend write your essay." And an INSEAD admissions officer supplied the countervailing honesty — "I don't think anyone is not using AI right now" — which is why almost nobody with a policy has banned the tool outright.

Imperial College Business School, in an official admissions post, explained what the offices are actually protecting: "When an application is heavily written or shaped by AI, we lose the ability to understand who you really are." That sentence is the whole issue. An application is an identification exercise. Anything that blurs the person — a ghostwriter, an over-editing parent, a model — defeats the file's purpose, whatever the policy says.

A former head of Wharton admissions described what readers now look for as the antidote: "a detail so specific it could only be true."

The detection question, answered honestly

Most applicants' first worry is being caught. The evidence says the worry is aimed at the wrong place.

AI detection software is unreliable in both directions, and the failure mode that should concern international applicants runs opposite to the one they fear. A Stanford study of seven commercial detectors found they flagged 61% of essays written by real, human, non-native English speakers as AI-generated; the same tools passed essays by native speakers at far higher rates. Detectors read low linguistic variance as machine writing, and careful second-language prose has low variance. Turnitin claims a false-positive rate under one percent; independent tests have found much higher, and universities including Vanderbilt have disabled its AI detector entirely. There is, as of this writing, no documented case of an M7 admission rescinded over AI detection.

So the software is not the risk. The risk is the human one the Cornell officer named: a reader who processes hundreds of essays a season recognizes the register, the balanced paragraphs, the borrowed vocabulary, the frictionless prose with nothing specific inside it. And the exposure does not end at submission. Essays are the script for interviews. A candidate whose spoken English sits two registers below their written English has revealed the ghostwriter without any detector involved, and this was equally true in the era when the ghostwriter was a paid consultant.

The schools' structural answer points the same direction. Kellogg expanded to five 60-second video questions, due within 96 hours of the deadline, and its admissions team was explicit that part of the point is that "you can't even use AI as a tool to help you answer those." Darden's new video essay allows ten seconds to read the prompt and 90 seconds to answer, in one take. Imperial states flatly that "reading from prepared answers or relying on AI tools during the recording is not permitted." Yale, on scripting your videos with AI: "don't." The written essay is shrinking and the live, unscripted sample of you is growing, precisely because the second cannot be delegated.

Where AI genuinely helps

None of this adds up to "avoid the tool." It adds up to a division of labor, and the division is not complicated.

AI is legitimately good at the work around the writing. Interrogation: handing a model your resume and asking for the twenty questions an admissions officer would press you on is a fast way to find the holes in your story. Organization: sorting a decade of experiences against a school's prompt types, checking whether a goals essay's logic actually closes. Editing: grammar and clarity passes on prose you wrote, which for non-native speakers is the highest-value, lowest-risk use available, and the one every disclosure policy comfortably covers.

What it cannot do is the excavation. The material that makes an essay work — the project where you first saw the gap you now want to close, the conversation that reordered your plans, the failure you are still slightly embarrassed by — exists only in your memory. A model cannot retrieve it; it can only pave over its absence with fluent generalities, which is exactly the surface readers have learned to recognize. The uniqueness an application needs is not manufactured at the keyboard. It already exists; the work is digging it out, and that work is questions, not prompts.

Two closing disciplines. First, whatever assistance you use, the disclosure boxes are the easiest integrity test of the season: answer them accurately, because the only way to lose on a disclosure question is to lie on it. Second, run the voice test before you submit: write a 200-word email to a friend on the same topic as your essay. If the essay does not sound like the email at sentence level, keep revising — not because a detector will flag it, but because an interview will.

• Draft alone, then bring the tool in. The defensible sequence is your draft first, AI for questions and edits after. The indefensible sequence is the reverse.
• Read each school's policy before opening a chatbot. A citation requirement at Kellogg, a footnote at London Business School, a checkbox at HBS: these take minutes to satisfy and are not worth improvising around.
• Spend the saved time on camera. Every hour AI saves you on prose is best reinvested in the one format the schools built specifically because AI cannot follow you into it.

The 2026–27 rules, read literally, describe a tool that may sharpen your sentences and question your logic, but may not speak for you. That is not a restriction on good applicants. It is a restriction on interchangeable ones, and this cycle, sounding like yourself has quietly become the scarcest asset in the pool.

Current as of August 21, 2026. AI policies changed mid-cycle last year and may again; verify each school's current language in its application before submitting.

Sources: Application instructions and policies published by Stanford GSB, Harvard Business School, Wharton, Columbia, Kellogg, London Business School, Michigan Ross, Duke Fuqua and NYU Stern; Clear Admit's "How to Use AI in Your MBA Application" admissions officer survey, August 2026, quoting Nicole Chen (Chicago Booth), Hannah Harlow (Cornell Johnson), Kate Botelho (Yale SOM), Jennifer Liu (INSEAD) and Judith Silverman Hodara (formerly Wharton); Poets&Quants, "MBA Applicant's Guide to AI in 2026" (May 2026); Imperial College admissions blog on responsible AI use (December 2025); the GradPilot business school AI policy index; Liang et al., "GPT detectors are biased against non-native English writers," Patterns.