Inside the Black Box: Detecting and Mitigating Algorithmic Bias across Racialized Groups in College Student-Success Prediction

Algorithmic bias in college success prediction is real and measurable: The study found that machine learning models used to predict college student success are systematically less accurate for racially minoritized students compared to white students, meaning these algorithms make worse predictions for students of color—potentially leading to unfair admissions and resource allocation decisions.

Historical data perpetuates past injustices: Because these predictive algorithms are trained on historical data, they inherit and amplify existing societal inequalities and racism embedded in that data, meaning colleges using these tools may inadvertently automate discrimination.

Standard bias-mitigation techniques are insufficient: The researchers tested leading technical approaches designed to reduce algorithmic bias but found that "common approaches to mitigating algorithmic bias are generally ineffective at eliminating disparities in prediction outcomes and accuracy between racialized groups"—suggesting that technical fixes alone cannot solve this problem.

Colleges need to reconsider their reliance on these tools: Given that these algorithms perform poorly for students of color and existing bias-reduction methods don't work well, institutions should carefully evaluate whether using student-success prediction algorithms is appropriate for high-stakes decisions like admissions and funding, or whether alternative approaches would be more equitable.