Every application season, I hear the same worry from applicants and their parents: a computer now reads your medical school application before a human ever does, and you somehow have to write for the algorithm. As a physician who served on a medical school admissions committee, I want to give you the accurate version of this story, because the rumor circulating on premed forums is far more sweeping than the reality.
A small number of medical schools have built machine learning models that perform the initial screen of applications. That first pass decides whether a faculty member ever opens your file. Most schools have not done this, and the two that clearly have kept human reviewers in control of every interview and admission decision. Here is what I know about which schools screen with AI, what those systems actually do, and how any of it should change the way you write your application, which is to say almost not at all.
When people say medical schools use AI to screen applicants, they are usually describing a much larger group than the evidence supports. Only two US medical schools have confirmed, published, live systems that sort applications. A few others are building tools they hope to pilot. Here is the current picture.
| Medical school | What the AI does | Status |
|---|---|---|
| NYU Grossman School of Medicine | A virtual faculty screener sorts each applicant into invite, hold, or reject for human review | Live |
| Zucker School of Medicine at Hofstra/Northwell | Acts as the first reader and recommends interview, further review, or reject | Live |
| George Washington University SMHS | A tool intended to help assess the full application | In development |
| University of Cincinnati COM | Plans to begin by screening essays, not yet deployed | In development |
| UC San Diego School of Medicine | Early discussions, no defined system | Exploratory |
The honest headline is that at the medical school level, only NYU Grossman and Zucker are confirmed live with published validation. Everything else is a plan, not a practice. As I explain below, the more consequential shift is happening one step later, in residency selection.
READ: AI in Medical School Admissions: Revolutionizing or Maintaining the Status Quo?
READ: Is Your Med School Consultant Using AI?
NYU built what its team calls a virtual faculty screener, a machine learning model designed to replicate how a human screener would triage an application into three buckets: invite for interview, hold for further review, or reject. The model was trained on a sample of more than 14,000 applications from 2013 to 2017, along with the actual decisions faculty screeners had made on those files, then tested prospectively in a subsequent cycle against the human reviews for the same applicants.
The motivation was volume. After NYU introduced full tuition scholarships, applications climbed sharply, and the school estimated that manual screening consumed more than 6,000 hours of faculty time each year. In the published validation, the model tracked human reviewers well on the clear invite and reject decisions and much less reliably on the ambiguous middle cases, which is exactly where you would expect an algorithm trained on human judgment to struggle. Marc Triola, MD, who led the work and is NYU's senior associate dean for medical education, has been candid that a model built from faculty decisions inherits the biases and flaws of those faculty along with their expertise. Final interview and admission decisions stay with the human committee. The research was published in Academic Medicine in 2023.
At Zucker, the AI is the literal first reader. It reviews the structured sections of roughly 5,000 applications a year and recommends whether each applicant should be interviewed, rejected, or routed to further human review, which is the largest group. According to the school, that step shrinks the pool of 5,000 down to somewhere between 1,500 and 2,000 files for committee review. Screeners on the admissions committee then work from that sorted pool to decide which 800 or so applicants receive interview offers.
Rona Woldenberg, MD, the school's associate dean for admissions, has framed the goal as consistency. A single tool applies the same standard to every file, which can reduce the variability that inevitably appears when many different humans screen thousands of applications on different days. To limit bias, the system tested for Zucker was deliberately built to withhold certain identifying details from the model, including the applicant's name, place of birth, and photo. The model was developed on more than 22,000 applications collected over five years and published in JAMIA Open in 2023.
I want to be fair about a limitation that the schools themselves acknowledge. When you train a model to reproduce past human decisions, you also teach it to reproduce whatever bias produced those decisions. Reducing the variability between reviewers is not the same as removing systemic bias, and the people building these tools are open about that tension. This is not a machine deciding your fate in isolation. It is a triage layer, and human committees still make every interview and admission decision.
Three schools appear in coverage of this topic but have nothing live. George Washington University's School of Medicine and Health Sciences is developing a tool meant to assess the entire application, and its admissions leaders have been direct that it is intended to support human review rather than replace it. As Ioannis Koutroulis, MD, PhD, the school's associate dean of MD admissions, put it, "This is not something that will replace human review." The scale explains the interest: GW receives roughly 13,000 applications a year, and each reviewer spends a large share of the season working through thousands of files. The University of Cincinnati College of Medicine plans to begin by assessing essays and faces a similar funnel, narrowing about 5,000 applicants to a class of roughly 180. UC San Diego has been mentioned only as being in early discussions, with no named tool. I would treat all three as intentions rather than practices, and I would not change a single word of your application because of them.
Some of the fear on this topic is aimed at the wrong targets. AMCAS and ERAS transmit your application as you submitted it. They are not ranking engines that sort or reject you. The Casper and AAMC PREview situational judgment tests are scored through human and psychometric methods, not an AI that decides your candidacy. When you hear that the whole application process has been handed to a machine, that is not accurate. The AI screening that genuinely exists is confined to the specific schools named above, plus the residency layer I turn to next.
While applicants argue about two medical schools, AI-assisted review has quietly become standard in residency selection. In July 2025, the AAMC and Thalamus made a tool called Cortex available at no cost to every residency and fellowship program that uses ERAS. Cortex uses machine learning, natural language processing, and optical character recognition to organize and normalize the material in an application so that program faculty can review it faster. It was first launched in 2020, has since been used across millions of applications, and reportedly cuts screening time by about half. Before the free rollout, it was piloted in orthopaedic surgery, physical medicine and rehabilitation, and urology.
The AAMC is explicit that ERAS itself does not use AI to sort or reject applicants, and that Cortex is meant to assist human reviewers under the organization's principles for responsible AI in selection. Still, if you are a current applicant, this is the version of the trend most likely to touch you, because it is already operating at scale. It is one more reason your ERAS materials need to be clearly organized and easy for a reviewer to parse quickly.
Here is the advice I give my own clients. At almost every medical school, a human still reads you first, so the idea that you must write for a bot is a distraction. Where AI does screen, it was trained on past human decisions, which means the signals that have always mattered still matter: academic rigor, a coherent narrative, genuine clinical exposure, and clear evidence of the qualities a given school values in its mission. There is no secret keyword that unlocks an algorithm, and anyone selling you one is selling you nothing.
I feel strongly about this next point. You should not be trying to game your application with buzzwords to please a screening tool, and I would push back hard on any advisor who tells you to. Stuffing an application with words like service, leadership, and resilience because you think a model is scanning for them does not make you a stronger candidate. It makes you a weaker one. These systems were built to reproduce how thoughtful faculty read applications, and thoughtful faculty see keyword padding for exactly what it is: a hollow application dressed up to look full. The applicants who write to impress a bot end up sounding like everyone else who is writing to impress a bot, which is the opposite of what gets you interviewed.
There is also a deeper reason not to do this. You are applying to become a physician, and the habit of performing for whatever is evaluating you is a poor one to build now, in a profession where honesty is not optional. I would rather you write one true, specific sentence about why medicine matters to you than a paragraph engineered to trip a filter.
What these systems actually reward is what strong applications have always shown, communicated in plain, well organized writing that a reader, human or machine, can follow without effort. Write in your own voice. Make your experiences legible. Let the substance of your candidacy do the work. That approach has never depended on who, or what, opens your file first, and it still does not.
A few do. Two US medical schools, NYU Grossman and the Zucker School of Medicine at Hofstra/Northwell, have live, published systems that use machine learning to perform an initial screen. A handful of others are developing similar tools. The great majority of schools still rely entirely on human screeners.
NYU Grossman and Zucker at Hofstra/Northwell are the two confirmed live examples. George Washington University and the University of Cincinnati are building tools, and UC San Diego has discussed doing so, but none of those three has a deployed, validated system.
No. AMCAS transmits your application to schools as you submitted it. It does not rank, sort, or reject applicants with AI. The same is true of ERAS at the residency level.
At the schools using these systems, the AI produces a recommendation as part of an initial screen. Human committees make the interview and admission decisions. The technology is a triage layer, not the final word.
No, and you should not try to. These models were trained on past human decisions, so they reward the same qualities strong applications have always shown. Padding your application with buzzwords to please a screening tool makes you a weaker candidate, not a stronger one, because it reads as hollow to the faculty the model is built to imitate. Write clearly, tell a coherent story, and let real experiences carry your candidacy. There is no keyword trick.
Yes, more so than at the medical school level. Since July 2025, the AAMC and Thalamus have offered a tool called Cortex free to all residency and fellowship programs that use ERAS. It helps faculty review applications more efficiently but is designed to assist human reviewers, not replace them.