Researchers at Handshake AI Research have found that AI tutors can provide learning gains for GRE (Graduate Record Examinations) preparation that are statistically equivalent to those provided by human experts.
NewsGemma 4 31BCurtis Northcutt
AI Tutoring Yields Learning Gains Equivalent to Human Experts for GRE Prep at a Fraction of the Cost
The study, led by Curtis Northcutt, compared the effectiveness and cost of three methods: interacting with AI tutors, receiving video-call instruction from human experts, and watching educational videos. The experiment involved 2,383 adults solving math and verbal GRE problems.
The results indicated that for math and verbal subjects combined, the average increase in correct answers was 13.9 points for the AI group and 15.6 points for the human expert group. While the human group showed a slightly higher mean increase, the 0.58-point difference fell within the research team's predefined "equivalent" range of ±4.09 points. However, when analyzed by subject, the equivalence was specifically confirmed for mathematics, while the verbal scores favored human instruction.
The cost efficiency of AI models was significantly higher. For instance, when comparing the cost per point of improvement, Gemma 4 31B was calculated at approximately $0.0052, compared to $4.81 for a human expert (assuming a human tutoring rate of $75 per hour). This makes Gemma 4 31B roughly 918 times more cost-effective than human instruction. Other AI models also proved more economical, with costs ranging from $0.024 to $1.40 per point of improvement.
The researchers noted that these findings specifically measure immediate learning gains following GRE preparation and emphasized that further research is needed to determine the long-term retention of the learned material.
Sources
- 「AI家庭教師と人間の専門家による指導の学習効果は同等」という研究結果、正答率の向上にかかる費用は918分の1 （GIGAZINE, 2026-09-26）
- arXiv