Source linked

Sierra Leone RCT de DeepMind: l'IA Tutoring fournit 1,2 à 1,7 ans de progrès en mathématiques en 8 semaines

Un essai randomisé contrôlé de 1 763 étudiants utilisant l'apprentissage guidé de Gemini a montré un gain de déviation standard de + 0,258 - équivalent à plus d'un an d'apprentissage typique - avec 91,4% des interactions construisant des concepts.

deepmindgeminifab aisierra leoneai educationrandomized controlled trial

0.258 standard deviations of math progress in eight weeks across 1,763 junior secondary students in Port Loko District, Sierra Leone. That number — from DeepMind's pre-registered randomized controlled trial of Guided Learning in Gemini — translates to roughly 1.2 to 1.7 years of typical learning gains. And when teachers integrated Gemini into half their lessons (targeting ~12 hours), the effect nearly doubled: 1.8 to 2.5 years of progress.

91.4% of Interactions Built Understanding, Not Just Answers

Over 113,000 student-Gemini exchanges were analyzed. In 91.4% of conversations, students used the tool to build conceptual understanding rather than fish for direct answers. Gemini responded by posing scaffolding questions in 76% of its messages, handing out direct solutions in only 2% of cases. That Socratic ratio is by design — LearnLM's pedagogical tuning prioritizes cognitive effort over convenience.

The behavioral shift accelerated over time. Skill-building queries rose from 68% in week one to 90% by week eight, while solution-seeking questions dropped from 25% to 10%. Students weren't just consuming answers; they were learning how to get there.

Teachers Remain the Center, Not the Bottleneck

Educators designed the lessons, set objectives, and facilitated classroom discussions. In focus groups, teachers reported using Gemini for lesson prep discovered new ways to explain fractions. Many described a shift from "lecturer" to "facilitator" — roaming the room while pairs of students worked through problems. DeepMind released a teacher training guide and an RCT playbook, co-created with Fab AI, for others running similar trials.

Engagement smashed the notorious "Five Percent Problem" (the typical voluntary edtech adoption rate): 69% of students met or exceeded usage targets. Students reported enjoying math more and voluntarily engaging beyond class hours.

The Achievement Gap Isn't Solved Yet

Students who entered the trial with stronger math skills benefited most. That gap is a known challenge, and DeepMind acknowledges the need for tools that deliver the strongest gains for the students who need it most. Future trials will expand to other countries and probe metacognition and relational intelligence.

DeepMind, in partnership with Fab AI, the Sierra Leone Ministry of Education, Google.org, and the Gates Foundation, is running additional pre-registered RCTs globally. The playbook and teacher training guide are live now — the evidence base for AI as a pedagogical partner just got a lot thicker.


Source: Measuring the impact of learning with AI in Sierra Leone and beyond
Domain: deepmind.google

Read original source ->

External source stays available while the OJO article and comment thread stay local.

Comments load interactively on the live page.