160k+ high school students will only graduate if a statistical model allows them to

algorithmic-fairnessml-biaseducationai-ethics

Abstraction: Critique of IB using statistical model to assign high-stakes graduation grades

Key points:

International Baccalaureate (IB) canceled 2020 exams for 160,000+ students across 144 countries and planned to assign grades via a statistical model using coursework grades, teacher predictions, and historical data
Seven identified methodological flaws: double jeopardy for low coursework scores, historical teacher bias against students of color, larger prediction errors for small schools (~25% higher error for 5-student vs. 300-student cohorts), measurement bias in under-resourced schools, inconsistent supplemental data, skewed grade distributions, and distribution shifts from teacher turnover
Even a 90% accurate model would incorrectly grade more students than all IB students combined in China, Germany, India, Singapore, and the UK
Fairness through unawareness fails: a model trained on test scores and school location predicted student majority race with higher accuracy than graduation rate, demonstrating the model was already race-aware
Three primary fairness criteria (group parity, equal error rates, equal precision) cannot be simultaneously satisfied — any model will violate at least two
Author argues for a process solution rather than a modeling solution when life outcomes of vulnerable populations are at stake

Connections: International Baccalaureate · Algorithmic Fairness · ML Bias · AI Ethics