Access Is Not Enough — Stanford Randomised Trials Find Students Barely Used an AI Reading Tutor, Even With Human Support

Researchers at Stanford (Carly Robinson, David Gormley, Ana Trindade Ribeiro and Susanna Loeb) ran two randomised controlled trials in 2024-2025 in two high-poverty US school districts, with 355 students in grades 1-5. An AI literacy tutoring platform (provider unnamed) was scheduled for two 30-minute sessions a week, in after-school homework time in one district and during class time in the other. Students used it either independently or with a human tutor focused on engagement rather than instruction, at a 2-5 to 1 ratio.
Uptake was low: just over 60% of students in one district and 53% in the other ever logged on, and average weekly use was 2.18 and 5.23 minutes against the provider's recommended 30. Human support increased weekly use by about 1 minute in one district and 4.4 minutes in the other (a 71-80% relative rise) but remained far below the recommended dose. The intervention did not improve reading achievement.
The authors stress that the tool never received an adequate dose, so the trials do not show that AI tutoring cannot work. Students who used it were less likely to receive special-education services and more likely to be higher achieving, a warning for equity. The practice is included as a cautionary case: provisioning AI tutors is not the same as getting students to use them.

Read the full analysis: https://nssa.stanford.edu/studies/access-not-enough-human-support-improves-engagement-ai-tutoring