Fairness Through Unawareness
What you'll learn: You'll understand the naive approach of excluding protected attributes from models, why it typically fails, and when it might still have value.
The Intuitive Idea
Fairness Through Unawareness is the simplest fairness approach: just don't give your model access to protected attributes like race, gender, or age. The logic seems sound—if the model can't "see" these features, it can't discriminate based on them, right?
Unfortunately, this approach almost always fails in practice.
Why It Fails: The Proxy Problem
The fatal flaw is proxy variables—features that are strongly correlated with protected attributes. Even if you remove gender from your dataset, features like:
- Job title ("nurse" vs "construction worker")
- Shopping patterns
- Zip code (correlated with race/ethnicity)
- Name patterns
...can allow the model to implicitly reconstruct the protected attribute and make biased decisions anyway. The model learns indirect pathways to discrimination.
Analogy: It's like removing someone's address from a form but leaving their zip code, neighborhood name, and nearby landmarks—you can still figure out where they live.
When It Might Be Useful
Despite its limitations, Fairness Through Unawareness can be a reasonable starting point when:
- Legal requirements mandate removing protected attributes
- No strong proxies exist in your feature set (rare in practice)
- Combined with other techniques as a preprocessing step
- You're establishing a baseline to measure against more sophisticated fairness interventions
Key Takeaway: Fairness Through Unawareness—simply removing protected attributes—is rarely sufficient because proxy variables allow models to indirectly learn biased patterns, though it may serve as a compliance or baseline step in broader fairness strategies.