Fairness Through Unawareness vs Awareness
You'll learn the fundamental tension between hiding sensitive attributes versus explicitly using them to achieve fair outcomes.
The Core Dilemma
When building fair machine learning systems, you face a critical choice: should you simply remove sensitive attributes like race, gender, or age from your model (fairness through unawareness), or should you explicitly account for them (fairness through awareness)?
Fairness through unawareness sounds intuitive—if the model can't see protected attributes, it can't discriminate, right? Unfortunately, this approach fails due to proxy variables (lesson 1889). Even without direct access to race, a model can learn patterns through zip codes, names, or education history that correlate strongly with protected classes.
Fairness through awareness takes the opposite approach: explicitly include sensitive attributes so you can measure and correct for disparate impact. This allows you to:
- Directly measure group-level fairness metrics
- Apply constraints ensuring equitable outcomes across groups
- Debug where bias enters your system
- Implement targeted interventions
The Paradox
Removing gender from a hiring model doesn't prevent discrimination if the model learns "attended women's college" predicts rejection. But including gender lets you enforce equal opportunity rates or test for disparate impact (lesson 1890).
Think of it like treating a medical condition: ignoring symptoms (unawareness) doesn't cure the disease. You need to diagnose with the information (awareness) to prescribe the right treatment.
Practical Reality
Most fairness-aware approaches temporarily use sensitive attributes during training or evaluation, then may remove them from production inference—getting benefits of both worlds.
Key Takeaway: Fairness through unawareness fails because proxy variables leak protected information; fairness through awareness explicitly measures and corrects bias, making it the more robust approach for achieving equitable outcomes.