Seminario 26/07 – Ismael Sanz (Universidad Rey Juan Carlos & London School of Economics) -Bridging Literacy Gaps: The Dose-Response Function of AI-Driven Personalised Learning

Información

School systems worldwide are investing heavily in AI-driven educational technology without guidance on a first-order question: how much should students use these tools? This paper estimates the dose-response function of DytectiveU, an AI-driven computer-assisted learning (CAL) programme deployed across 263 primary schools in the Madrid region from 2019 to 2023. Using granular server-side log data for over 33,000 students and more than 907,000 sessions, we exploit quasi-random variation in the timing of programme activation across school-grade-year cohorts as an instrument for student usage. Each additional session raises in-platform literacy proficiency by 0.60 percentage points, more than double the OLS estimate of 0.27; the corresponding intention-to-treat magnitude is 2.3 percentage points (0.10 standard deviations) per month of earlier activation — a direct policy-relevant counterpart to the LATE. The empirical distribution of usage is heavily skewed: the median student completes only 9 sessions, against a recommended dose of 64. Combined with evidence that the production function exhibits diminishing marginal returns — robust to selection on unobservables, sign-consistent across every subgroup we examine, and confirmed by a recentered influence function analysis showing the ITT declining monotonically from 3.9 pp per month at the 10th percentile of proficiency to 0.9 pp per month at the 90th — this pattern implies that most students operate on the steep portion of the dose-response curve and that onboarding interventions would be distributionally progressive. Subsample estimates are suggestive of larger returns for students in grades 1–3 than for grades 4–6, consistent with developmental sensitive periods in literacy acquisition, though a pooled interactive specification does not reject equality of returns and we therefore treat this as a descriptive rather than a formally identified gradient. Relative to Bettinger et al. (2023), who provide the first experimental evidence on diminishing returns to CAL by comparing two discrete dosage levels, our contribution is to estimate the continuous dose-response function from a scaled government implementation, allowing us to characterise both the shape of the in-platform production function and where students actually operate on it.