Sensitive Attribute

Sensitive attributes derive their status from anti-discrimination laws and regulatory frameworks. In the US, the Equal Credit Opportunity Act (ECOA) protects attributes like race, color, religion, national origin, sex, marital status, and age in lending decisions. The EU's General Data Protection Regulation (GDPR) classifies these as special categories of personal data under Article 9, prohibiting processing unless specific conditions are met. Key protected classes include:

• Race and ethnic origin
• Gender and sexual orientation
• Age and disability status
• Religious or philosophical beliefs
• Genetic and biometric data

Even when a sensitive attribute is explicitly excluded from training data, models can infer it through proxy variables—seemingly neutral features that correlate strongly with protected characteristics. For example, ZIP code often correlates with race due to historical housing segregation, and browser type or device model can correlate with income level. This phenomenon, known as redlining by proxy, makes fairness-aware engineering challenging. Detection requires:

• Correlation analysis between features and protected attributes
• Causal graph construction to identify backdoor paths
• Adversarial testing to measure inferability of sensitive data

Sensitive attributes are often unobservable or uncollectable in production systems due to privacy regulations, consent requirements, or organizational policy. This creates a fundamental tension: fairness evaluation requires knowing group membership, but collecting that data may itself be prohibited. Common approaches to this paradox include:

• Differential privacy mechanisms that add calibrated noise
• Third-party escrow services that hold sensitive labels for auditing
• Bayesian improved surname geocoding (BISG) to probabilistically estimate race from name and location
• Fairness through unawareness (widely recognized as insufficient)

Evaluating fairness across single dimensions can mask intersectional harms—disparities that only emerge at the intersection of multiple sensitive attributes. A model may appear fair for women overall and fair for Black individuals overall, yet systematically disadvantage Black women specifically. This concept, rooted in Kimberlé Crenshaw's legal scholarship, demands:

• Multi-dimensional subgroup analysis beyond top-line metrics
• Worst-case subgroup performance monitoring
• Granular evaluation across all combinatorial slices of protected attributes
• Recognition that the most vulnerable subgroups are often the smallest

The classification of what constitutes a sensitive attribute is not static—it evolves with societal norms and legal precedent. Attributes like genetic information gained protected status only after the Genetic Information Nondiscrimination Act (GINA) of 2008. Similarly, political affiliation and social media behavior are emerging as contested categories. Engineering implications include:

• Designing governance frameworks that accommodate new protected classes
• Building model monitoring systems with configurable fairness dimensions
• Maintaining audit trails that can retroactively evaluate historical decisions against newly recognized sensitive attributes

Not all correlations with sensitive attributes constitute unfairness. A critical distinction exists between statistical association and causal discrimination. For example, in medical diagnosis, biological sex may be causally relevant to disease risk and should inform predictions. The challenge is distinguishing legitimate causal pathways from illegitimate discriminatory ones. Techniques include:

• Causal fairness criteria like counterfactual fairness
• Path-specific effects analysis to isolate discriminatory causal channels
• Structural causal models (SCMs) to encode domain knowledge about legitimate vs. illegitimate dependencies

A protected characteristic encompassing an individual's racial identity, ethnic origin, or skin color. In AI systems, race is a prohibited basis for decision-making under civil rights laws such as the U.S. Civil Rights Act and the EU's Racial Equality Directive.

• Proxy risk: Features like ZIP code, surname, or purchasing patterns can inadvertently encode racial information
• Redlining prevention: Fair lending models must explicitly test for racial disparities in credit decisions
• Audit requirement: Regular disparate impact analysis comparing approval rates across racial groups

A protected attribute covering sex, gender identity, and gender expression. Algorithmic discrimination on this basis is prohibited under laws including the Equal Credit Opportunity Act and GDPR Article 9.

• Representation harms: LLMs and recommendation systems can perpetuate occupational stereotypes (e.g., associating nursing with women)
• Non-binary exclusion: Systems requiring binary male/female inputs systematically exclude non-binary individuals
• Pay equity: Compensation algorithms must be audited for gender-based wage gaps after controlling for legitimate factors

A protected characteristic under the Age Discrimination in Employment Act (ADEA) and similar global regulations. Age-based algorithmic discrimination often manifests in hiring, credit, and insurance underwriting.

• Digital redlining: Younger users may receive predatory financial product offers while older users are excluded from digital services
• Cold start amplification: Age-correlated behavioral patterns can cause recommendation systems to limit content diversity for older demographics
• Intersectional risk: Age combined with gender creates compounded bias vectors requiring multi-dimensional fairness testing

Protected under the Americans with Disabilities Act (ADA) and the UN Convention on the Rights of Persons with Disabilities. AI systems must provide equitable access and outcomes regardless of physical, sensory, or cognitive disability.

• Accessibility bias: Computer vision models trained predominantly on non-disabled bodies may fail to recognize or appropriately serve users with mobility aids
• Speech recognition gaps: Voice interfaces often exhibit higher error rates for users with speech impairments, creating service exclusion
• Employment screening: Automated resume parsers must not penalize employment gaps related to disability leave

A protected characteristic under Title VII of the Civil Rights Act and Article 9 of the GDPR, which classifies religious beliefs as special category data requiring explicit consent for processing.

• Dietary and product filtering: Recommendation systems must avoid inferring religious affiliation from purchase patterns to prevent discriminatory pricing or exclusion
• Temporal patterns: Prayer timing, religious holidays, and observance patterns create behavioral signals that models may inadvertently exploit as proxies
• Content moderation: Automated systems must not disproportionately flag or suppress religious expression while permitting secular equivalents

A sensitive characteristic protected under GDPR Article 9 and increasingly under state and national anti-discrimination laws. Inference or use of sexual orientation in algorithmic decisions is explicitly prohibited in many jurisdictions.

• Inference from behavior: Social graph analysis, media consumption, and purchase history can reveal orientation with high accuracy without user knowledge
• Outing risk: Personalization systems that surface LGBTQ+ content may inadvertently disclose orientation to third parties viewing a shared screen
• Differential pricing: Travel and hospitality algorithms have been documented charging premium rates in LGBTQ+ neighborhoods