📄 Full text: EdWorkingPaper 26-1451 · local
Robinson, Gormley, Ribeiro & Loeb (2026) ran two RCTs showing that AI tutoring's binding constraint is take-up, not capability: despite dedicated session time, nearly half of students never used the platform and users averaged only 2–5 minutes per week. An in-person engagement tutor (not direct instruction) raised usage by 1–4 minutes/week and engagement by 71–80% — but dosage stayed far below the level needed for reading gains, and achievement did not improve.
The two RCTs
- Districts A and B, elementary students, AI literacy (reading) platform; treatment = in-person tutor focused on engagement (10-min check-in + 15 min platform + 2–5 min reflection), control = platform alone with the full ~30 min
- Pre-registered outcomes: minutes/week (usage) and stories read/week (engagement); spring ELA as achievement
- Interventions lasted 14–31 weeks
Take-up is the story
- Only 60.7% / 53.3% of control students ever used the platform
- Average weekly usage: 2.18 / 5.23 minutes; even users-only averages were 13.2 / 25.8 minutes
- Students used the platform in only 4–5 of 14–31 weeks
- Platform users skewed higher-achieving and less likely to receive special-education services — an equity flag: the students who need adaptive support most are least likely to engage
What human support did
- Usage: +1 min/week (A), +4.4 min/week (B) — significant, but a drop against the ~30 min/week the provider recommends for measurable reading gains
- Engagement: +0.20 stories/week (A, +71%) and +0.92 stories/week (B, +80%)
- Total added dosage: ~22 minutes (A) and ~98 minutes (B) across the entire intervention
- No reading achievement gains in either district (negative, non-significant) — consistent with the achieved dosage
- Strong site-level variation: some sites saw little effect, others meaningful gains — local implementation conditions matter
Connections to the wiki
- Direct evidence on the intelligent-tutoring take-up problem: system quality is worthless without student engagement
- Extends the equity analysis of AI tutoring: access gaps compound with engagement gaps (special-ed and lower-achieving students used less)
- The engagement-tutor mechanism (relationship-building, check-ins, reflection) resonates with human-in-the-loop and the relational care arguments of care-full-feedback-genai
- Contrasts with learner-ai-interaction-patterns-oop: even when usage is measured, the dosage threshold for achievement effects is often missed
- Supports the "design for learning" stance of beyond-detection-authentic-assessment-ai-2025: engagement must be designed, not assumed
Related Pages
- intelligent-tutoring — the platform technology whose take-up is the binding constraint
- equity — who uses (and who is left out of) AI learning tools
- engagement-metrics — usage and story-completion as the measured outcomes
- k-12 — elementary school context
- ai-literacy — the platform's domain
- student-experience — motivation and participation
- human-in-the-loop — the tutor-supported engagement architecture
- care-full-feedback-genai — relational support as the engagement lever
- learner-ai-interaction-patterns-oop — measuring real usage patterns
- beyond-detection-authentic-assessment-ai-2025 — engagement must be designed, not assumed
Sources
- Robinson, C. D., Gormley, D., Trindade Ribeiro, A., & Loeb, S. (2026). Access is Not Enough: Human Support Improves Engagement with AI Tutoring. EdWorkingPaper No. 26-1451, Annenberg Institute at Brown University. DOI