When an AI explains a difficult problem, the explanation can feel clear. That feeling is useful, but it leaves a question unanswered: can you solve a different problem yourself?
For students, parents and people learning new skills at work, that is a better way to judge a tutor than the polish of its answers.
The new finding: immediate GRE gains
The September 30 revision of StudentBench reports results from 2,383 participants completing 2,469 GRE study sessions. After one hour, pooled AI tutoring produced immediate unaided learning gains statistically equivalent to expert human tutoring within predefined margins. Compared with the no-tutoring group, its adjusted gain was 6.15 percentage points. [1]
STUDENTBENCH · PREPRINT
A narrow result, not a teacher-replacement verdict
2,383 participants · One-hour sessions · Immediate unaided GRE tests.
Equivalence held for the combined comparison and Quantitative section, but not for Verbal separately. It means similarity within a statistical margin, not identical performance.
The work was funded by Handshake AI and is a preprint. Participants were English-reading adults recruited for a paid study. Retention over months was not measured. The control watched unrelated educational videos; there was no practice-only group. Consequently, this experiment does not isolate tutoring’s benefit over independently practicing GRE problems. [1]
The counterpoint: help can become a crutch
A separate randomized high-school mathematics experiment in Turkey, published in PNAS in June 2025, tested nearly 1,000 students. Unrestricted GPT-4 assistance increased practice performance by 48% relative to control, but subsequent unaided exam performance was 17% lower. A version designed to guide students with hints largely prevented that harm; it did not establish a positive unaided exam effect. [2]
HIGH-SCHOOL MATH · GPT BASE ONLY
Better practice. Weaker unaided performance.
Relative differences versus the study’s control group, not percentage-point changes. Bar lengths show absolute magnitude on a 0–50% scale; color and signs indicate direction. Different assessments, not a before-and-after score trajectory. Source: Bastani et al. [2].
These studies do not contradict each other. They test different tools, learners and educational tasks. Together they support evaluating the learner’s later performance, rather than assuming every kind of assistance has the same effect.
Known, uncertain and speculation
KNOWN
Design matters.
The tested tutoring approaches produced different outcomes. A polished answer is an incomplete measure of educational value.
UNCERTAIN
Will it last?
These findings cannot establish durable learning across ages, languages and subjects, or the performance of every current product.
SPECULATION
Broader access.
Affordable tutoring could expand learning opportunities. Whether that becomes lasting educational progress remains an open question.
A practical learning check
Our recommendation: use a short cycle that makes your own reasoning visible. This is an editorial exercise, not a separately validated intervention.
- Attempt: try a problem before asking for assistance.
- Ask for a hint: explain where you got stuck and request one next step.
- Explain: describe why your solution works in your own words.
- Close the tool: solve a new problem independently, then revisit it another day.
If you are learning software testing, ask AI to explain a failing assertion, then write a test for a different scenario yourself. Being able to inspect, explain and correct the result gives you a more useful signal than simply obtaining working code.
A useful learning tool should leave you with more capability to bring to the next problem.
Primary sources and evidence notes
- Northcutt et al. — StudentBench · September 30, 2026 revision · Preprint; Handshake AI funded. Released code and data.
- Bastani et al. — Generative AI without guardrails can harm learning · June 25, 2025 · Peer-reviewed randomized experiment. August correction concerns an author affiliation, not these results.