Last week’s issue walked through how to match a student to the right Common App prompt before a word gets written, including the closing structure that has become a specific tell in AI-polished essays. This issue picks up immediately after that decision. Once a prompt is chosen and a draft begins, what actually protects a student’s voice through the process that follows is a separate question, and the answer has less to do with any single tell and more to do with the research now available on what AI use is doing to essays across an entire applicant pool.
Run the same kind of human written paragraph through a detector twice. Attribute one to a native English speaker and one to a non native speaker, and the false positive rate jumps from about 5% to 61%. That finding comes from a peer reviewed Stanford study, and the gap it describes is the real problem with treating AI detection as a solved question in admissions. The tools are not reliably catching artificial writing. They are catching writing that does not sound like a fluent native speaker, and for an IEC working with English language learners or international applicants, that distinction is most of the job.
This newsletter grows from your shares. Send it to a fellow colleague today!
In today’s issue (free analysis):
What Common App’s fraud policy actually defines as AI misuse, and why the enforcement mechanism behind it looks more like an honor code than a scanner
What four years of essay data at one selective college reveal about who is using AI to write, and why those students are more likely to be rejected even when their essays read as polished
Why Duke stopped assigning essays a numerical score, and what that decision signals about where other admissions offices are likely headed
What IECA’s own members told their association is the largest concern in the profession right now, ahead of financial aid or enrollment cliffs
For paid subscribers:
An intake to final draft sequence that documents a student’s actual writing process without turning every session into paperwork
Three voice calibration exercises that establish what a student’s sentences actually sound like before any tool touches the page
The exact conversation to have when a student arrives with a draft that already has AI in it, and the one to have when a parent assumes you are using AI yourself
How to adjust essay coaching for a student who writes in a second language, given what the research says about which writers get falsely flagged
Paid subscribers get the full analysis every issue, including the practitioner frameworks in the section below the paywall. If this newsletter helps you stay ahead in your work, a paid subscription is one of the better professional investments you can make. Many readers expense it as professional development.
Already a paid subscriber? Thank you - you’re the reason this exists.
What Common App actually polices, and what it doesn’t
Common App’s own fraud policy, last updated to address generative AI directly, defines fraud as submitting plagiarized material or intentionally misrepresenting as one’s own original work either another person’s language and ideas or the substantive content or output of an artificial intelligence platform, technology, or algorithm. All applicants submitting through Common App agree to this definition of fraud, regardless of which member institution they apply to.
What the policy does not contain is a percentage threshold, a defined detection method, or a stated reliance on automated scanning software. The policy’s own description of its enforcement process is complaint driven. A registered user, a member college, or a third party can report suspected fraud, the accused student is notified and given an opportunity to respond, and Common App investigates from there. Common App does not publicly describe any routine scanning of essays, and no evidence currently indicates that such screening occurs at the platform level.
Individual colleges layer their own statements on top of this baseline, and the range is wide. Brown’s admissions office states that AI use in application content is not permitted under any circumstances, with an explicit exception for spelling and grammar review. Brown separately conducts random application verification as part of its broader integrity process, though it does not describe this as an AI specific measure. Yale’s admissions office publishes a dedicated AI Policy Statement drawing the same line: AI for grammar review or general topic suggestions at the start of the writing process is fine, AI generated content in a personal statement may result in admission revocation or expulsion. Plenty of colleges have published nothing at all. For those schools, the operative standard is still the Common App attestation every applicant signs, certifying that the application is their own work, factually true, and honestly presented.
The practical upshot for an IEC is that the question worth coaching toward is not whether a particular tool would flag a given paragraph. No verified evidence from any of the policies above describes a school running every submitted essay through a detector as a matter of course. The standard that actually governs the outcome is closer to a credibility judgment made by a human reader and, if a question is ever raised, an investigation triggered by a complaint. That is a different kind of risk than a software score, and it should be coached differently.
A nuance worth keeping in mind: Admissions offices rarely disclose their internal review processes, so the absence of a public detector policy should not be interpreted as proof that no screening tools are used.
Four years of essays, and what changed after 2022
A research team from Cornell and Carnegie Mellon spent four years inside the essay files of one unnamed selective institution, comparing tens of thousands of submissions from before generative AI tools existed through several cycles after. Two findings from that work matter directly to how an IEC should think about essay coaching.
Disclaimer: Because the research examines a single institution and has not yet undergone peer review, the findings should be interpreted cautiously.
The first is about who uses AI and what it does to their odds. A 2026 working paper from researchers at Cornell and Carnegie Mellon examining one selective institution found evidence that applicants who received an application fee waiver, used as a proxy for lower income, were more likely than other applicants to use AI in their essays. Among the group of applicants who used AI, lower income applicants were still more likely to be rejected than higher income applicants who also used it. Jinsook Lee, the study’s lead author and a doctoral candidate at Cornell, has suggested this gap likely reflects access. Higher income applicants are more likely to have counselors, teachers, or paid tools that help them use AI well, while a student relying on a free tier tool gets a noticeably weaker output. This research suggests AI may not function as the equalizer it is sometimes assumed to be. It widens an existing gap in who has good writing support.
The second finding is about what AI use is doing to the essays themselves. The researchers measured how similar essays are to one another, a property they call homogenization, and found it increased significantly after generative AI tools became available, with the largest increase concentrated among lower income applicants and applicants who were ultimately rejected. AJ Alvero, a Cornell sociologist and co-author on the study, has described the concern plainly. The personal statement exists to let an applicant show the specific, individual texture of their own life. If a growing share of essays are converging on the same template, that opportunity is quietly disappearing for the students who relied on it most.
There is a third, more granular finding worth knowing for coaching purposes. Lee has observed that large language models have a habit of forcing identity markers into a sentence where they do not belong, an example being an essay opening with a phrase like “as an Asian woman” attached to a sentence that has nothing to do with that identity. It reads as the model trying to perform specificity it does not actually have access to.
A separate, earlier Cornell analysis, led by associate professor Rene Kizilcec, compared AI generated essays directly against 30,000 human written essays from before ChatGPT existed. The AI essays were highly generic and struggled to build a genuinely individual narrative, and giving the AI specific personal details about the applicant sometimes made the output sound more mechanical rather than less, as the model worked those details in as keywords rather than lived material. The same research team trained a classifier that distinguished the AI essays from the human ones with near perfect accuracy, within their controlled research dataset.
That last result is worth a caution of its own. The classifier was built and tested on a known, labeled set of essays where the researchers already knew which were AI generated and which were not. That is a meaningfully easier task than what an admissions reader faces with a live application of unknown origin, where a student may have used AI for part of a draft, revised it heavily by hand, or not used it at all. A research result showing that a purpose built classifier can separate two known categories in a controlled dataset is not the same claim as saying admissions offices are reliably catching AI writing in the field. The honest summary of the evidence is that AI generated essays tend to read as generic to a trained eye, and that a researcher can build a tool that confirms this in a controlled setting. Whether any given admissions office is running anything like that tool on live applications is, with the data available right now, unconfirmed.
Duke’s answer was not a better detector
In early 2024, Duke stopped assigning a numerical score to applicants’ essays and standardized test scores, a change confirmed independently by Inside Higher Ed and reported in detail by the Duke Chronicle. Dean of Undergraduate Admissions Christoph Guttentag explained the reasoning directly to the Chronicle. The decision was driven largely by the rise of AI generated writing and, separately, by concern over heavily ghostwritten essays produced by paid consultants. “Essays are very much part of our understanding of the applicant,” Guttentag wrote. “We’re just no longer assuming that the essay is an accurate reflection of the student’s actual writing ability.”
Duke kept the essay in the file. What changed is what the essay is being asked to prove. It no longer contributes to a numerical score alongside curriculum strength, grades, recommendations, and extracurriculars. It is read now for the purpose Guttentag described to his own student newspaper, to help the office understand the applicant as a person rather than as a set of achievements.
Duke's change occurred relatively early in the current wave of admissions related AI policy discussions. But the direction is informative regardless of how many other offices follow it explicitly. No institution examined here publicly emphasizes AI detectors as its primary solution. The schools with the clearest public policies are managing the problem through attestation and consequence, not technology, and Duke went a step further and changed what the essay is used to evaluate. Either response points an IEC toward the same coaching priority. An essay’s job is shifting away from proving a student can write well in isolation and toward giving a specific reader something they could not have gotten anywhere else in the file. That is a craft standard, not a detection standard, and it is the one worth coaching toward regardless of what any individual school’s policy says.
What is actually keeping IECA’s members up
In May, more than one thousand independent educational consultants gathered in Baltimore for IECA’s 50th annual conference. Stephanie Simpson, who became the association’s CEO last December, told Inside Higher Ed that a recent survey of IECA’s own membership produced an unambiguous result. According to IECA leadership, AI emerged as one of the association's most significant concerns. This was both, in how to use it ethically in their own practices and in how to guide students who are already using it.
Lisa Carlton, IECA’s incoming president and a consultant who began her career nearly twenty years ago working with neurodivergent students, described the texture of that anxiety in practical terms rather than abstract ones. She has watched colleagues experiment carefully with AI as a research and brainstorming aid, crafting prompts deliberately so a student gets a structured starting point rather than an unfiltered flood of information from the open internet. She also offered a specific reframe for how to respond when a student brings in something they found through AI or social media that the IEC did not generate and might not fully trust: rather than dismissing it, ask what felt right to the student about it, and offer to add context rather than override their instinct. That posture, treating the student’s own engagement with the tool as a starting point to build from rather than a mistake to correct, is the through line of how IECA’s leadership is currently describing good practice.
Carlton also raised a caution that connects directly back to the Cornell and Carnegie Mellon equity finding. As some institutions experiment with alternative formats like video essays or portfolios in response to concerns about written essays, she flagged that these formats can quietly disadvantage students who do not have the same access to time, equipment, or polish that a wealthier applicant might. The same access gap that produced a lower income essay penalty in the homogenization study shows up again whenever a new format is introduced without accounting for who can actually produce it well.
Knowing what the policy actually says and what the research actually shows is the diagnosis. The coaching session itself is a separate problem, the sequence that builds a documented writing process before a draft exists, the exercises that surface a student’s actual voice, and the language for the two conversations every IEC eventually has with a student or a parent. That work is below.
Higher Ed Insights is reader-supported. If you are reading this as a free subscriber and you work in college counseling, financial aid, enrollment management, or high school education, the paid section is written directly for you.