Disparate Treatment

The defining characteristic of disparate treatment is the direct ingestion of a protected attribute (e.g., race, gender, age) as a model feature. When a credit-scoring algorithm explicitly assigns lower scores based on a protected class variable, it constitutes disparate treatment. This is distinct from fairness through unawareness, where the attribute is merely removed but may still be proxied by other features.

Legally and technically, disparate treatment hinges on mens rea—a guilty mind or purposeful action. In an ML context, this means a model architect or business stakeholder knowingly designed a system to treat groups differently. Evidence includes:

• Documented feature selection that explicitly includes protected class status
• Business rules that encode different thresholds for different groups
• Internal communications revealing a motive to disadvantage a protected class

Disparate treatment and disparate impact are often conflated but are legally distinct doctrines:

• Disparate Treatment: Intentional discrimination. A smoking gun. The policy is discriminatory on its face.
• Disparate Impact: A facially neutral policy that disproportionately harms a protected group, regardless of intent. Measured by the 80% rule or adverse impact ratio. A model that excludes a protected attribute can still produce disparate impact through proxy discrimination.

Auditing for disparate treatment involves code review and feature provenance analysis rather than purely statistical tests. Key steps include:

• Feature inventory: Cataloging every input variable and its legal classification
• Data lineage tracing: Verifying that no protected attribute leaks into training data through joins or derived features
• Causal graph construction: Mapping whether any feature is a deterministic function of a protected attribute If a model's decision function f(x) changes when only the protected attribute is altered, disparate treatment is confirmed.

In the United States, disparate treatment in automated decisions is prohibited under statutes including:

• Title VII of the Civil Rights Act of 1964: Bars employment discrimination based on race, color, religion, sex, or national origin
• Equal Credit Opportunity Act (ECOA): Prohibits credit discrimination on the basis of race, color, religion, national origin, sex, marital status, or age
• EU AI Act: Classifies AI systems that manipulate or exploit vulnerabilities based on protected characteristics as unacceptable risk, effectively banning them

The primary remediation for disparate treatment is pre-processing elimination of the protected attribute and its direct proxies. This is more surgical than post-processing or in-processing fairness interventions:

• Schema-level enforcement: Automated checks that prevent protected class columns from entering feature stores
• Disparate impact testing on the sanitized model to catch residual proxy effects
• Counterfactual fairness verification to ensure predictions are invariant to protected attribute perturbations Note that simple removal (fairness through unawareness) is insufficient without addressing correlated features.

A form of algorithmic bias where a non-protected feature serves as a stand-in for a protected attribute. This allows disparate impact to occur indirectly, even when protected attributes are excluded.

• Redlining via zip codes is a classic example
• ML models can discover proxies automatically through correlation
• Defeats fairness through unawareness strategies

The systematic and repeatable error in an ML model that creates unfair outcomes, privileging one arbitrary group over another. Bias can originate from:

• Representation bias: Skewed sampling in training data
• Historical bias: Existing prejudices reflected in data
• Measurement bias: Flawed labels or feature selection
• Disparate treatment is one manifestation of algorithmic bias

A legally or ethically recognized characteristic that must not be used as a basis for discriminatory decision-making. Common protected attributes include:

• Race, color, national origin
• Sex, gender identity, sexual orientation
• Age (40+ under ADEA)
• Disability status
• Religion

Disparate treatment occurs when these attributes are explicitly used to deny opportunities.

A causal definition of fairness where a prediction is fair if it remains the same in the actual world and a counterfactual world where the individual belonged to a different demographic group. This directly tests for disparate treatment by asking: "Would the decision have been different if the person were of a different race?"

• Requires a structural causal model
• Captures both direct and indirect discrimination

A naive intervention where a model is simply denied direct access to protected attributes. This is widely considered ineffective because:

• Proxy variables (zip code, surname, browsing history) carry the same information
• Models can reconstruct protected attributes from correlated features
• Does not prevent disparate treatment via proxies

Often called "fairness through blindness" or "anti-classification."