What Happened
A new study says an AI tutor can match what an expert human tutor does for GRE test-takers, at a fraction of the price. Released on September 23 as a preprint titled "StudentBench: AI and human tutoring yield equivalent GRE learning gains," the paper describes a public evaluation platform built to answer a narrow but consequential question: when a student sits down with a large language model instead of a person, do they actually learn as much? Researchers Curtis Northcutt, Inaara Hasmani, Kevin Feng, Trevor Khangi, Andreas Plesner and Jonas Mueller put 2,383 participants through one of three conditions — AI tutoring, human tutoring, or no tutoring at all — then measured how much their scores moved on Quantitative and Verbal GRE questions. The result: "AI tutoring is statistically equivalent to expert human tutoring for GRE learning gains," with a p-value of .015, a threshold researchers use to say a result is unlikely to be chance.
The platform behind the numbers, StudentBench, logged more than 175,000 student-AI messages over the course of the study, giving the team a data set large enough to break results down by GRE domain rather than treat the exam as one lump score. In five of the seven GRE content areas tested, the researchers found the best-performing AI tutor actually surpassed the human tutor's average gains, not merely tied it. The study ran as a timed, one-hour session: a student took a pre-test, practiced with their assigned tutor on problems drawn from what they missed, then sat a different post-test to see how much ground they had made up in that single hour.
A second, separate study had expert human tutors run 2,028 pairwise rubric comparisons, judging AI-generated lesson plans and practice problems against each other rather than against a human's work. That exercise let the authors separate AI tutors along five dimensions — lesson planning, practice-problem creation, conversational pedagogy, cost and student engagement — instead of collapsing everything into a single pass-fail grade. The full data set and the tutoring platform itself are public at studentbench.org, and the paper does not name which specific AI models were tested or publish a per-model leaderboard.
Why It Matters
The headline number in the paper isn't really about test scores — it's about price. The authors calculate that one AI tutor reached learning gains equivalent to a human tutor's at 918 times lower cost: 0.52 cents versus $4.81, per percentage point of score improvement. That gap is the kind of figure that reorders a market rather than nudging it. Private GRE tutoring in the United States commonly runs well into three figures per hour, and graduate-school hopefuls already spend heavily on prep courses, practice tests and one-on-one coaching in the run-up to an exam that still shapes admissions and, in some fields, funding decisions.
StudentBench lands in a beat that outlets tracking AI in education describe as thin but growing quickly this year, following moves like OpenAI's teen-focused AI guidance and a wider push by ed-tech companies to fold tutoring features into consumer chat products. What sets this paper apart is that it is a controlled comparison rather than a product announcement: a no-tutoring control group, matched pre- and post-tests, and a peer-reviewable methodology rather than a vendor's internal benchmark. For a field that has mostly relied on usage statistics and anecdote to argue AI tutoring works, a study built specifically to test that claim against a human baseline carries more weight than another feature launch.
The pressure point lands hardest on the test-prep industry itself. Firms built around in-person and live-online tutoring, from large national chains to independent coaches, price their hourly rates around the cost of skilled human labor. If a machine can produce comparable score gains for a fraction of a cent, the calculus for what a family or a university tutoring budget should pay for changes, and pricing built around scarce human time becomes harder to defend on learning outcomes alone.
Reaction
Coverage of the paper has focused less on whether AI tutoring works — plenty of prior research has gestured at that — and more on what the cost gap means for people who currently pay for tutoring. One industry newsletter's editorial take put it bluntly: if the 918-times figure holds up under independent replication, "test-prep pricing and university tutoring budgets are the first line items to revisit," with human tutors shifting "from default option to premium tier" rather than disappearing outright. That framing treats the result as a repricing event for an existing service, not a wholesale replacement of tutors, and it's a distinction the paper's own authors seem careful about: they highlight where AI tutors won on five of seven GRE domains, but they don't claim victory across the board, and they stop short of recommending AI tutoring replace all human instruction.
What's Next
Because StudentBench is a public platform rather than a one-off experiment, the team behind it can keep collecting data as more students use it, and the authors flag one especially practical finding for anyone trying to build on this work: for Quantitative GRE sessions, faster AI replies were linked to students sending more messages, more of those messages containing correct practice work, and more correct practice linked to bigger score gains — a chain of correlations the paper reports as statistically significant. That points toward response latency, not just model capability, as a lever platform builders can pull to improve outcomes.
The obvious next question is whether the GRE result generalizes: to the SAT, LSAT and other high-stakes exams; to K-12 subjects with different pedagogical demands; and to learners who aren't already comfortable typing questions into a chat interface. The paper's authors don't make that leap themselves, and independent replication — ideally by researchers with no stake in either AI tutoring products or the traditional tutoring industry — is the natural check on a single-study result, however large its sample. The platform's public availability at studentbench.org at least means outside teams can attempt exactly that.
Closing Thoughts
What's notable about StudentBench isn't that an AI produced good tutoring — that claim has circulated for a couple of years now. It's the shape of the evidence: a randomized comparison against both a human baseline and a no-tutoring control, run at a scale most classroom studies never reach, with the cost difference spelled out in cents rather than vague claims of "affordability." That combination is what separates a research finding from a sales pitch, and it's why the paper is being read as a marker for the tutoring market rather than just another benchmark score.
None of that settles what happens to the people whose livelihoods depend on hourly tutoring rates, and the study doesn't pretend to. What it does is narrow the argument from "does AI tutoring work" to a more answerable one: at what price does a family, a school or a university decide human time is still worth paying for, once a much cheaper alternative can show it gets students to the same place.
한글 요약
2026년 9월 23일 공개된 새 연구 논문 “StudentBench”에 따르면, AI 튜터가 GRE(미국 대학원 입학시험) 학습 성과에서 숙련된 인간 튜터와 통계적으로 동등한 수준을 보였다. 커티스 노스컷 등 6명의 연구진은 2,383명의 참가자를 AI 튜터링, 인간 튜터링, 무튜터링 세 그룹으로 나눠 1시간 세션 전후 점수 변화를 측정했으며, 그 결과 p값 .015로 AI와 인간 튜터링이 통계적으로 동등하다는 결론을 얻었다. 7개 GRE 영역 중 5개 영역에서는 최고 성능 AI 튜터가 평균적으로 인간 튜터를 능가하기도 했다.
이번 연구에서 가장 주목받는 수치는 비용 차이다. 한 AI 튜터는 인간 튜터와 동등한 학습 성과를 점수 1퍼센트포인트당 0.52센트로 달성했는데, 이는 인간 튜터 비용(4.81달러)의 918분의 1에 불과하다. 업계 뉴스레터는 이 격차가 독립적으로 재현된다면 "입시 컨설팅 요금과 대학 튜터링 예산이 가장 먼저 재검토해야 할 항목"이 될 것이며, 인간 튜터는 "기본 선택지에서 프리미엄 등급으로" 자리를 옮기게 될 것이라고 평가했다. 다만 연구진은 AI가 모든 영역에서 인간을 앞섰다고 주장하지는 않았으며, 구체적으로 어떤 AI 모델을 사용했는지는 논문에 명시하지 않았다.
연구팀이 공개한 StudentBench 플랫폼은 누구나 studentbench.org에서 이용할 수 있으며, 앞으로 GRE 외에 다른 시험이나 학습 영역으로 검증 범위를 넓힐 수 있는 기반이 될 전망이다. 연구진은 정량 영역 세션에서 AI의 응답 속도가 빠를수록 학생이 더 많이 연습하고 더 큰 점수 향상으로 이어지는 상관관계도 발견했다. 다만 이번 결과가 SAT나 LSAT 등 다른 고부담 시험, 혹은 K-12 교육 현장에도 그대로 적용될지는 아직 확인되지 않았으며, 독립적인 후속 검증이 필요하다는 것이 연구진과 업계 관계자들의 공통된 지적이다.