The research behind ThinkHumanly
Most edtech makes claims that don't hold up. We've built our framework explicitly on the research that does replicate — and explicitly avoided the work that doesn't.
Theoretical foundation
Vygotsky's Zone of Proximal Development
Activities should land in the gap between ‘can do alone’ and ‘can do with help.’ Murray & Arroyo (2002) operationalized this for adaptive systems.
Self-Determination Theory (Ryan & Deci 2017)
Three psychological needs drive intrinsic motivation: autonomy, competence, and relatedness. Meta-analyses (Wang et al. 2024) show autonomy support has effect sizes around g=1.14.
Diamond's Executive Function model (2013)
Three core EFs — working memory, inhibitory control, cognitive flexibility — are foundational to academic and life outcomes. Our resilience and self-direction dimensions tap these.
CASEL 5 competencies (2020)
Self-awareness, self-management, social awareness, relationship skills, responsible decision-making. Strong meta-analytic evidence (Durlak 2011, Taylor 2017) for outcomes.
Csikszentmihalyi's flow (Shernoff et al. 2003)
Engagement peaks when challenge meets skill. We monitor effort signals to keep activities in the productive zone.
Effect sizes that matter
Active learning
Freeman et al. (2014) PNAS: students in active learning environments are 1.5× less likely to fail STEM courses; effect size +0.47 SD on exam scores.
Project-based learning
Chen & Yang (2019) meta-analysis: d+=0.71 effect on student academic achievement.
AI-personalized STEM learning
2025 meta-analyses: g=0.42–0.53 effect across studies.
Intelligent tutoring systems
2025 meta-analysis: g=0.27 vs. traditional instruction.
What we explicitly disavow
Learning styles matching
Pashler et al. (2008) and follow-ups: no replicable evidence that matching teaching to claimed ‘learning styles’ improves outcomes. We treat our 7th dimension as engagement preference (a starting bias for hybrid ratio), NOT a learning style.
Growth mindset miracle claims
Macnamara (2022) meta-analysis: average effect d̄=0.08, far smaller than popular claims. Sisk (2018) ~1% variance. We respect growth-oriented language without promising outcomes.
Cognitive diagnosis from short interactions
A 10-minute screen activity cannot reliably assess intelligence, attention disorders, or learning disabilities. We never claim it can.
Voice or biometric child data
Beyond research validity concerns, this creates DPDP/COPPA/GDPR-K regulatory risk and ethical issues. We do not collect it.
Our scoring methodology
Parent intake (14 questions) provides the primary signal at 67% weight. The optional child activity contributes 33% as a corroborating signal, capped at ±15 points so a short session can't flip a dimension. Confidence intervals are reported transparently. Equal-weighted item averaging is our defensible v1; we'll refine weights as we gather convergent validity data from teacher reports (target r ≥ 0.3-0.4) in school pilots.
The team behind this
We're building ThinkHumanly with input from learning scientists, classroom teachers, child psychologists, and the parents of the children we serve. Methodology questions, citations, or concerns: research@thinkhumanly.co