Particle.news

Custom AI Tutor Doubles Learning Gains in Harvard–MIT Trial

Researchers say teacher‑informed instructional design explains the gains; careful implementation is required for wider use.

Overview

  • A randomized controlled trial reported in September 2026 found that 194 students using a bespoke AI tutor called PS2 Pal showed median learning gains roughly twice those recorded after a best‑practice active learning physics lesson.
  • Students who used the tutor also reported higher engagement and motivation and spent less time on the material than peers in the classroom condition.
  • The tested system was a custom implementation built on the GPT API with pre‑vetted prompts, structured feedback, verified content, and self‑paced practice rather than an open‑ended public chatbot.
  • Researchers and a Brookings center report attribute the results to deliberate instructional design and educator co‑design, and they warn that AI can harm learning if it simply supplies answers or is deployed without those supports.
  • School districts are piloting AI for personalization and attendance early‑warning uses even as some jurisdictions, including New York City, impose limits to study equity, vendor practices, and classroom impacts before broader scaling.