AI tutors match expert humans on GRE learning gains at 918x lower cost
arampell · x · 2026-09-25
- Curtis Northcutt's team (Cleanlab/MIT) released the StudentBench paper (arXiv:2609.28470), testing 2,383 real students: pre-test → 1 hour with AI, human, or no tutor → post-test. No LLM judges; gains are post minus pre.
- AI tutoring is statistically equivalent to expert human tutoring for GRE learning gains (p=.015); the best AI tutor beat human experts on average in 5 of 7 GRE domains.
- Cost: $75/hour for a human expert vs $0.07/hour for AI — 918x cheaper (per-problem $0.0052 vs $4.81, p=.044).
- A second study had experts run 2,028 pairwise rubric evaluations of AI-generated lesson plans and practice problems, separating AI tutors on lesson planning, problem creation, conversational pedagogy, cost, and engagement.
- The platform logged 175,000+ student-AI messages. Motivation: $1 trillion/year goes to making machines better; some should go to making humans better.
Related event: AI Tutors Match Human Experts at 918x Lower Cost, Study Finds(2 posts)→
More from AGI Musings
- Gary Marcus on CNBC to Talk Meta, OpenAI, and Why 'We All Might Be Screwed' — GaryMarcus · 2026-09-25
- "Alignment isn't a panacea" — researcher argues for hardening non-AI systems instead — gandamu_ml · 2026-09-25
- Kai-Fu Lee: open vs closed models and data sovereignty decide AI's fate beyond the two superpowers — kaifulee · 2026-09-25
- What Jack Butcher's AI performance art is really telling us to do — WhatTheLJW · 2026-09-25
- Predicting a domino run of AI regulation ending in a state-mandated duopoly — max_paperclips · 2026-09-25
- 'Want' is clearer than 'feel': a debate over whether models can think — JeffLadish · 2026-09-25