Virtual Tutoring with Computer-Assisted Learning: An Experiment in Take-Up and Learning

TWiK (Tutoring With Khan Academy) — a low-cost online after-school tutoring layer in which one-on-one human virtual tutors supervise struggling Grade 4–8 students working through the Khan Academy practice their teacher already assigned. A two-year randomized trial with the Toronto District School Board finds the binding constraint is take-up and sustained participation, not tutor quality: only 45% of assigned students reached a first session in Year 1, but collapsing enrollment to a single step raised first-session take-up to 83% in Year 2 — while weekly attendance stayed near 40% because students attended intermittently rather than dropping out. The offer raised practice by ~10 minutes a week and produced small, imprecise gains (ITT 0.055 SD on math assessment; ~0.08 SD on report-card marks) plus improved math attitudes, effects that grew with take-up. The core lesson for AI tutoring and computer-assisted learning alike: the platform supplies content, but someone must get the student to practice regularly.

Key Findings

• Take-up is the primary obstacle. In Year 1 only ~45% of teacher-nominated, randomly assigned students attended even one virtual tutoring session; families were lost at every step of a long chain (notice invitation → express interest → create an account on a third-party platform → get matched → show up). Students were lost at each handoff.
• Simple design changes nearly doubled first-session take-up. In Year 2, inviting families directly through their own teacher for a specific weekly time, handling enrollment and scheduling internally, and setting up a recurring session on the family's behalf raised first-session take-up from ~45% to 83% — evidence that taking up an offer is highly responsive to reducing friction.
• Solving entry is not solving participation. Even among students who started, weekly attendance remained near two-fifths; "attrition" was mostly intermittent absence (missing some weeks, returning others), not dropout. Practice among attendees rose only ~17 minutes a week.
• Modest, coherent gains. The intent-to-treat effect on a short achievement test covering assigned Khan Academy topics was 0.055 SD (95% CI [−0.09, 0.20]) — positive but too imprecise to separate null from a small effect. Report-card marks rose ~0.08 SD (p < 0.10), and treated students reported enjoying math more and finding the platform more helpful.
• Effects scale with take-up. Gains grew between Year 1 and Year 2 as first-session take-up rose, consistent with a program whose modest dose produces modest effects that strengthen when more students actually engage.

Connected Concepts

• Intelligent Tutoring
• Human In The Loop AI
• Edtech Platform
• Math Education
• K 12
• Help Seeking
• Student Engagement
• Adaptive Learning
• Learning Gains
• Equity In AI Education

Connected Articles

• Elevate GenAI Virtual Tutors — GenAI virtual tutors
• Access Not Enough AI Tutoring 2026 — Access is not enough for AI tutoring
• One Click Away Khanmigo Two Year School Experiment 2026 — One Click Away: Khanmigo in a two-year school experiment
• AI Tutoring Quality K12 Methodologies 2026 — Improving AI tutoring quality in K-12
• AI Availability Student Motivation — AI availability and student motivation

Citation

Oreopoulos, P., Dong, R., & Low, N. (2026). Virtual tutoring with computer-assisted learning: An experiment in take-up and learning (NBER Working Paper No. 35622). National Bureau of Economic Research.