StudentBench Study Finds AI Tutoring Matches Human Tutoring on GRE Learning Gains

A study published as a preprint on September 23, 2026 found that AI tutoring produced GRE learning gains statistically equivalent to expert human tutoring, at a fraction of the cost. The research, called “StudentBench: AI and human tutoring yield equivalent GRE learning gains,” was conducted by Curtis Northcutt, Inaara Hasmani, Kevin Feng, Trevor Khangi, Andreas Plesner and Jonas Mueller, and funded by Handshake AI Research, the AI-training-data arm of the jobs platform Handshake. Northcutt is also co-founder and chief executive of the data-quality company Cleanlab. The AI tutoring learning gains finding rests on an equivalence test with a p-value of .015, meaning the researchers could not statistically distinguish the average gains of AI-tutored students from those tutored by human experts.

The study built StudentBench, a public evaluation platform, and logged more than 175,000 student-AI interactions to support its analysis. Its first experiment measured learning outcomes directly; a second experiment had expert human tutors rate the quality of AI-generated teaching materials.

How the study measured AI tutoring learning gains

Researchers recruited 2,383 adult participants and randomly assigned them to one of three conditions: AI tutoring, human tutoring or an unassisted control group using standard study resources. Participants completed a pretest, one hour of tutoring or study time on GRE quantitative or verbal reasoning material, and a posttest. The sample split into 1,291 quantitative-reasoning sessions and 1,178 verbal-reasoning sessions, with 2,139 sessions using AI tutors, 140 using human tutors and 190 in the control group. A second study had 51 expert human tutors complete 2,028 pairwise comparisons judging AI-generated lesson plans and practice problems against benchmarks, drawing on 381 student pretests.

Pooling results across models, the paper reports statistical equivalence between AI and human tutoring within a ±0.25 standard-deviation bound (p = .015), and again within a tighter ±0.20 standard-deviation bound (p = .023). Looked at by subject, AI tutors’ best-performing models exceeded the human tutor’s average learning gains in three of four quantitative domains, while human tutors retained the advantage across all three verbal domains tested. The researchers also found that faster AI response times were associated with higher student engagement and larger learning gains in quantitative sessions, a relationship significant at p < .002.

Cost differences between AI and human tutoring

The study’s starkest figure is cost. Measured in dollars spent per percentage point of test-score improvement, the cheapest AI tutor tested, Gemma 4 31B, cost $0.0052, compared with $4.81 for a human tutor, a 918-fold difference. Other AI models fell in between, reflecting a tradeoff between higher-end model cost and tutoring quality.

Source: Northcutt et al., “StudentBench: AI and human tutoring yield equivalent GRE learning gains,” arXiv, September 23, 2026.

What the researchers say the results do not show

The paper’s authors frame the equivalence finding narrowly: it applies to GRE preparation over a one-hour session, not to tutoring generally, and human tutors still led on verbal-reasoning gains and on the quality ratings applied in the second study. Stanford research grading AI tutoring tools has previously found mixed performance across AI tutoring products, and a separate Stanford study finding students barely use AI tutoring when it is offered showed that access alone does not guarantee engagement. A two-year trial covered by Winss Solutions, the Khanmigo AI tutoring study that found small maths gains, reported more modest effect sizes than the StudentBench comparison, underscoring how outcomes vary by tutoring product, subject and study design.

Background

Handshake AI Research is the AI-training-data division of Handshake, a US jobs and recruiting platform for college students, which has expanded into producing human-verified training data and benchmarks for AI models. StudentBench, the open evaluation platform built for this research, remains publicly available at studentbench.org, where users can take a free AI-tutored GRE practice session structured around a pretest, one hour of tutoring and a posttest. The authors say the underlying data and code needed to reproduce every figure and table in the paper are published on GitHub. The study adds a directly comparative, cost-quantified data point to a growing body of research on AI tutoring’s classroom and test-preparation value, following months of mixed evidence from other AI tutoring evaluations that Winss Solutions has tracked this year.

Sources: arXiv (Northcutt et al., “StudentBench: AI and human tutoring yield equivalent GRE learning gains”); StudentBench (Handshake AI Research); AI Weekly

Featured image: photo by Katerina Holmes on Pexels (free Pexels license).

Become a Sponsor

Our website is the heart of the mission of WINSS – it’s where we share updates, publish research, highlight community impact, and connect with supporters around the world. To keep this essential platform running, updated, and accessible, we rely on the generosity of you, who believe in our work.

We offer the option to sponsor monthly, or just once choosing the amount of your choice. If you run a company, please contact us via info@winssolutions.org.

• donate monthly
• donate once only

Select a Donation Option (USD)

$5.00 (p/m) $10.00 (p/m) $20.00 (p/m) Other (p/m)

Enter Donation Amount (USD)

Figen Sekin

I specialize in sustainability education, curriculum co-creation, and early-stage project strategy. At WINSS, I craft articles on sustainability, transformative AI, and related topics. When I’m not writing, you’ll find me chasing the perfect sushi roll, exploring cities around the globe, or unwinding with my dog Puffy — the world’s most loyal sidekick.