bettermarks — Real-World Effectiveness of an AI-Adaptive Maths Platform Across German State School Systems

bettermarks is a German-developed AI-adaptive mathematics platform that generates individualised practice problems and step-by-step feedback based on each student's error patterns. Following a 2009 pilot study by the ZNL Transfer Center for Neuroscience and Learning at the University of Ulm (34 schools, one semester, a significant positive effect on fractions performance), the platform scaled to statewide licences in at least 9 of Germany's 16 federal states, with the company reporting roughly 500,000 student users and 500+ schools as of August 2025.

Its most rigorous independent evaluation to date is a peer-reviewed 2026 field study by Schaaf, Rolfes, Nagy and Heinze (Frontiers in Education), which followed 587 grade 8-9 students across 60 classes in 25 socioeconomically disadvantaged Schleswig-Holstein schools from a pretest (Sept-Oct 2024) to a posttest (Feb-Mar 2025). An earlier October 2017 impact evaluation by Perera and Aboal (Centro de Investigaciones Económicas, commissioned by Uruguay's Plan Ceibal) tracked 2,143 primary-school students across 237 schools using the same platform, locally called PAM.

Average system use in the German study was just 9.2 minutes/week (4.6 in school, 4.5 at home). Average learning gains across the full sample were near zero, but students who used the system more than their classmates during school hours showed a small, statistically significant individual-level benefit (β=.13, p<.01); home use and class-level (aggregate) usage effects were non-significant. In Uruguay, the earlier study found a robust 0.20-standard-deviation positive effect on mathematics learning gains, with the effect growing larger for students completing more exercises — a clear dose-response pattern.

Read together, bettermarks can help — particularly students who actually use it consistently, especially in a structured, in-school setting — but average, unconditional effects are small or null, and impact depends heavily on dosage and implementation, not the software alone.