What Bias Looks Like in ML Models
You'll learn: How algorithmic bias manifests concretely across real-world domains and why it matters for ML practitioners.
The Reality of Algorithmic Bias
Bias in ML isn't abstract—it shows up as measurably unfair outcomes. When models encode societal prejudices or dataset imbalances, the consequences are real: denied loans, missed diagnoses, wrongful arrests, and perpetuated discrimination.
Concrete Examples Across Domains
Criminal Justice: Recidivism prediction models have flagged Black defendants as "high risk" at twice the rate of white defendants with similar histories, while underpredicting risk for white defendants. The model's predictions perpetuate historical arrest patterns rather than actual reoffending.
Healthcare: Algorithms allocating medical resources have systematically assigned lower risk scores to Black patients than white patients with identical health conditions. The bias stemmed from using healthcare spending as a proxy for need—but Black patients historically receive less care due to access barriers.
Hiring: Resume-screening models trained on past hiring decisions have learned to downrank candidates from women's colleges or with "foreign-sounding" names, reproducing historical discrimination patterns in new decisions.
Credit & Finance: Loan approval models may deny credit to qualified applicants from minority neighborhoods, even when not explicitly using race, because the model learned correlations between ZIP codes and default rates shaped by redlining history.
Face Recognition: Models achieve 99%+ accuracy on light-skinned males but error rates over 30% for dark-skinned females, resulting in misidentification and false arrests.
How Bias Manifests
Bias appears through disparate error rates (higher false positive/negative rates for protected groups), different prediction distributions (systematically higher/lower scores), and outcome disparities (unequal access to opportunities or resources).
Key Takeaway: Algorithmic bias isn't hypothetical—it manifests as measurable disparities in predictions and outcomes across protected groups, often amplifying historical inequities embedded in training data or problem formulation.