Protected Attributes and Sensitive Features
What you'll learn: You'll understand what protected attributes are, how they differ from proxy variables, and why their use raises critical legal and ethical questions in ML systems.
What Are Protected Attributes?
Protected attributes (also called sensitive features) are characteristics of individuals that are legally or ethically protected from discrimination. Common examples include:
- Race or ethnicity
- Gender or sex
- Age
- Religion
- Disability status
- National origin
In many jurisdictions, decisions based on these attributes are illegal in contexts like hiring, lending, housing, and healthcare.
Protected Attributes vs Proxy Variables
Here's the tricky part: even if you remove protected attributes from your training data, your model can still learn to discriminate. How? Through proxy variables—seemingly innocent features that correlate strongly with protected attributes.
Analogy: Imagine you remove "gender" from a hiring model but keep "years of career break." Since career breaks disproportionately affect women (due to maternity leave), the model indirectly learns gender patterns. The career break variable becomes a proxy for gender.
Common proxies include:
- ZIP codes (proxies for race/income)
- Names (proxies for gender/ethnicity)
- Educational institutions (proxies for socioeconomic status)
Legal and Ethical Considerations
Legal: In the U.S., laws like the Civil Rights Act and Equal Credit Opportunity Act prohibit discrimination. In the EU, GDPR restricts processing of sensitive personal data.
Ethical: Even when legal, using protected attributes or their proxies can perpetuate societal inequities, harm marginalized groups, and erode trust in AI systems.
The paradox: Sometimes you need protected attributes to measure and mitigate bias—but using them in predictions may be illegal or unethical.
Key Takeaway: Protected attributes are legally/ethically sensitive characteristics like race and gender; removing them isn't enough because proxy variables can encode the same information, requiring careful analysis of all features and their relationships.