Picture a premed student stuck on an enzyme kinetics problem late at night. She puts it into a chatbot, and within seconds she has a step-by-step explanation, five practice questions on the same concept, and a stack of flashcards for the morning. Before ChatGPT launched in November 2022, that kind of help usually meant paying a tutor or high priced study materials. Now it’s free or close to it.
If tools like that make students better at studying, the Medical College Admission Test is a good place to look for proof. It’s a high-stakes, closed-book exam, and its scores are built to mean the same thing from one year to the next.
So Premed Plug went looking in public data from the Association of American Medical Colleges, which administers the MCAT. What it found: the national average hasn’t gone up. Across all recorded exam results, the average total score was 500.6 in 2023–2025, compared with 501.6 in 2019–2021, before ChatGPT launched.
That doesn’t mean AI is failing students. The AAMC’s tables don’t record who used AI, so they can’t tell us whether it helped or hurt the people who did. They only tell us the average hasn’t moved.
It isn’t because students are ignoring the tools. We found no published survey measuring AI use specifically among MCAT examinees, and the AAMC’s own surveys of examinees and medical students haven’t asked about it. But the broader numbers are hard to miss. In a 2025 Inside Higher Ed survey of 1,047 students at 166 U.S. colleges, 85% said they’d used generative AI for coursework in the past year; half had used it like a tutor, and 46% had used it to study for exams or quizzes. Medical students got there early. In a May 2023 survey of 415 students from 28 medical schools, 52% had already used ChatGPT for coursework.
The national average hasn’t risen
The broadest measure covers every recorded exam result, including repeat attempts and people who never go on to apply to medical school. Each year, the AAMC publishes percentile tables built from every score in the three most recent testing years. Comparing two windows that don’t overlap, 2019 through 2021 and 2023 through 2025, sets results from before ChatGPT’s release beside results from after AI tools went mainstream. They’re different groups of test-takers, of course, and the earlier window includes pandemic-era testing.
The average total score across all exam results was 501.6 in the earlier window and 500.6 in the later one, a one-point decline on a scale that runs from 472 to 528. The spread of scores also widened slightly, with the standard deviation rising from 10.8 to 11.2 points.
A one-point dip is small, less than a tenth of a standard deviation, and nobody should read it as proof that AI made students worse. But students clearly have not figured out how to turn it into a gain.
Top scores didn’t climb either. About 12% of exam results landed at 515 or above in 2019–2021, compared with about 11% in 2023–2025, according to the AAMC’s percentile tables. At the other end, the share at 490 or below rose from about 17% to about 20%. These are shares of all results combined, so they say nothing about how any particular group, such as strong students who used AI, would have scored otherwise.
The same flat pattern holds for students who went on to apply to medical school. The average MCAT score among applicants to U.S. MD programs was 505.6 in the 2018-19 application cycle, 506.5 in 2022-23, and 506.3 in 2025-26, according to AAMC data. One wrinkle: application cycles aren’t testing dates. Most medical schools accept MCAT scores up to three years old, so each cycle mixes scores from several earlier testing years and trails any change in how students prepare.
Most of that rise came before ChatGPT existed. The enrolled-student average had already reached 511.9 by 2021-22, then held roughly steady before ticking up in 2025.
Selection may be part of it. Applications rose 5.3% in 2025, to 54,699, according to the AAMC. But schools also enrolled a record class of 23,440, so it’s hard to say how much tougher the competition for each seat really got, or how much that explains.
The pool of test-takers changed, too. The number of recorded MCAT exam results, which include repeat attempts, rose roughly 10% between the 2019–2021 and 2023–2025 testing windows, to more than 305,000. A shift in who takes the exam can move averages on its own, in either direction.
Could AI make MCAT prep cheaper and more accessible?
There’s another way AI could change MCAT prep that score averages miss entirely: cost. Traditional help doesn’t come cheap. Kaplan lists its self-paced MCAT course at $1,599, its 515+ course at $3,599, and private tutoring packages starting at $3,599, though promotions vary. In the AAMC’s 2025 Post-MCAT Questionnaire, 30.1% of respondents reported participating in a structured prep course.
Students have also shifted toward cheaper, self-directed prep. In the 2025 questionnaire, 55.4% of respondents reported using other free MCAT prep resources, and 50.8% used free online flashcard generators. In 2020, those shares were 39.3% and 37.3%, according to Student Doctor Network’s analysis of earlier AAMC reports. AI tools, many with free or low-cost versions, could extend that trend. Established prep companies are adopting AI too: a comparison on Kaplan’s website lists an AI tutor in the MCAT prep offered by Kaplan, Blueprint, The Princeton Review, and UWorld.
But cheaper doesn’t automatically mean better. Nancy Romano of DAT Destroyer, a test-prep company and longtime Student Doctor Network sponsor, told the site that students are experimenting with AI for flashcards, quick explanations, and practice drills, but that its accuracy and depth are often inconsistent compared with guides written and edited by subject experts. The AAMC’s own official prep materials also cost money, though students who qualify for its Fee Assistance Program can get reduced testing fees and free access to them. Whether AI ends up narrowing the gap between students who can pay for help and those who can’t is still an open question.
Research suggests tool design matters
We couldn’t find a published study linking students’ AI use to their MCAT scores. Two recent experiments in other subjects come closest. Their headlines point in opposite directions, but they land in the same place.
In a field experiment published in the Proceedings of the National Academy of Sciences in 2025, researchers from the University of Pennsylvania gave nearly 1,000 high school math students in Turkey access to AI tutors built on GPT-4 during practice sessions. Students using a standard, unrestricted chatbot did far better on practice problems while they had it. But on an exam taken without AI, they scored 17% lower than students who never had access.
A second version, built to give hints instead of handing over answers, avoided that drop, but it didn’t produce a statistically detectable improvement on the unassisted exam either.
A 2025 randomized trial at Harvard, published in Scientific Reports, found a more encouraging result. Among 194 undergraduates in an introductory physics course, students learned significantly more, in less time, from a carefully designed AI tutor than from an in-class active-learning session. That tutor was built to guide students through problems rather than give them answers. The study measured learning right after each lesson, so it doesn’t show long-term retention or any effect on a test like the MCAT.
Put the two together and a pattern emerges. AI that hands students answers seems to backfire once it’s taken away, while AI built to make students do the work has done better, at least in the short term.
That’s exactly the situation on the MCAT, a closed-book exam of about seven and a half hours in which most questions ask students to apply concepts to passages describing experiments or data. Whatever tools students study with, the test measures what they can do without them.
What the data can’t tell us yet
The biggest gap in all of this is that the AAMC’s score data doesn’t identify which test-takers used AI, or how much. Timing adds another limitation. More capable models and purpose-built study tools arrived mostly in 2024 and 2025, so the latest testing window combines results from early and later stages of adoption, and application-cycle data adds a further lag.
The way the exam is scored helps, but only so far. The AAMC equates scores across test dates, so a given score reflects the same level of performance no matter which version of the exam a student took. That accounts for differences in exam difficulty, not differences in who is taking it, so changes in the test-taking population could mask gains or losses among AI users. The next percentile update, due in May 2027, will add 2026 results.
Use AI as a tutor, not an answer key
None of this is a reason to swear off AI. The national averages haven’t budged, but they can’t tell any single student whether it’s working for them. What the research does suggest is that the best AI help leaves students able to work without it. In practice, that means trying a question before asking for an explanation, asking for practice questions instead of summaries, checking answers against trusted prep materials, and taking full-length timed practice exams with no AI at all. On test day, it’s just the student and the exam.


