---
source_url: https://doi.org/10.26300/pz7p-p388
ingested: 2026-08-03
sha256: 87884fc922f83a9897c9b717fccc598a801d1442c22789a306b0fcb6b5e66bc8
---

# Access is Not Enough: Human Support Improves Engagement with AI Tutoring

Carly D. Robinson, David Gormley, Ana Trindade Ribeiro, Susanna Loeb (Stanford University / SCALE Initiative). EdWorkingPaper No. 26-1451, Annenberg Institute at Brown University, June 2026. DOI: 10.26300/pz7p-p388.

## Design

Two randomized controlled trials (Districts A and B) in elementary schools. Students assigned to use an AI literacy platform (reading/stories) either independently (control) or with an in-person tutor focused on engagement — not direct instruction (treatment). Tutor sessions: 10-min check-in + 15 min platform work + 2-5 min reflection; control students had the full ~30 min for platform work. Stratified randomization by site × grade/classroom. Interventions lasted 14-31 weeks. Preregistered outcomes: platform usage (minutes/week) and engagement (stories read/week); BOY assessments as controls, spring ELA (EOY) as achievement outcome.

## Take-up problem (control)

- Only **60.7% (District A) and 53.3% (District B) of control students EVER used the platform** despite dedicated session time.
- Average weekly usage: **2.18 min (A), 5.23 min (B)**; among users only: 13.2 (A), 25.8 (B) min/week.
- Students used the platform during only 4-5 weeks on average across 14-31 week interventions.
- Platform users were less likely to receive special education services and more likely to be higher achieving.

## Treatment effects (ITT)

- No effect on likelihood of ever using the platform (effects not driven by one-time log-ins).
- **Usage: +1 min/week (A), +4.4 min/week (B)** — statistically significant but far below the ~30 min/week recommended for measurable reading gains.
- **Engagement: +0.20 stories/week (A) = +71%** (control mean 0.28); **+0.92 stories/week (B) = +80%** (control mean 1.14).
- Total added dosage: ~22 min (A) and ~98 min (B) over the whole intervention.
- **No reading achievement gains** in either district (spring ELA: negative, non-significant point estimates) — consistent with achieved dosage being far below threshold.
- Site-level heterogeneity: some sites showed little/no effect, others meaningful increases — local implementation conditions matter.

## Implications

- Access alone is insufficient; implementation and human support matter for meaningful engagement with AI-based learning tools.
- AI tutoring's promise of scaling personalized instruction depends on student take-up, which can be the binding constraint.
- Even with human support, engagement gains can remain below the dosage needed for achievement effects.