A protected attribute is a personal characteristic, such as race, sex, age, religion or disability, that law or policy shields from discrimination, and that fairness analysis in AI ethics uses to define the groups whose outcomes must be compared. In machine learning it is also called a sensitive attribute, and it is usually written as A to distinguish it from the ordinary input features X and the outcome Y.
The concept links anti-discrimination law with statistical evaluation. A model may never receive a protected attribute as input and still treat groups differently, so practitioners treat the attribute as a reference variable for auditing rather than simply as a column to delete.
Where protected attributes come from
Protected attributes are defined by legal and ethical frameworks, not by data science. Which characteristics count depends on the jurisdiction and the domain, such as employment, credit, housing or public services, so the same model can face different protected lists in different markets.
The table below summarizes common sources and the characteristics they name. The lists are simplified; each instrument has its own scope, exceptions and case law.
| Source | Domain | Examples of protected characteristics |
|---|---|---|
| US Title VII of the Civil Rights Act | Employment | Race, color, religion, sex, national origin |
| US Age Discrimination in Employment Act | Employment | Age 40 and older |
| US Equal Credit Opportunity Act | Lending | Race, sex, marital status, age, receipt of public assistance |
| UK Equality Act 2010 | Employment, services, education | Nine characteristics including disability and sexual orientation |
| EU Charter of Fundamental Rights, Article 21 | All EU law | Sex, ethnic origin, religion, disability, age, sexual orientation and others |
Data protection law adds a related but distinct category. The GDPR lists special categories of personal data, such as health and ethnic origin, whose processing is restricted. The overlap creates a known tension: auditing for bias may require collecting the very data that privacy rules limit.
How protected attributes enter fairness metrics
Fairness metrics use the protected attribute to split predictions into groups and compare them. Without knowing A for at least an evaluation sample, almost no group fairness metric can be computed, which is why the attribute is central even in systems that never use it for decisions.
Each metric conditions on the protected attribute in a different way, and the choice encodes a normative view of what equal treatment means.
| Criterion | Condition across groups of A | Typical concern addressed |
|---|---|---|
| Demographic parity | Equal positive prediction rates | Unequal access to opportunities |
| Equalized odds | Equal true and false positive rates | Unequal error burdens |
| Equal opportunity | Equal true positive rates | Qualified people being missed |
| Predictive parity | Equal precision of positive predictions | Scores meaning different things per group |
| Disparate impact ratio | Selection rate ratio, often checked against 0.8 | Adverse impact under the four-fifths rule |
These criteria are mutually incompatible in general: when base rates differ between groups, a classifier that is not perfect cannot satisfy calibration and equalized error rates at once. Choosing a metric is therefore a policy decision, not a purely technical one.
Why removing the attribute is not enough
Deleting the protected attribute from training data, an approach known as fairness through unawareness, rarely prevents discrimination. Other features often act as proxies that encode the same information, so the model can reconstruct group membership indirectly.
This effect is called proxy discrimination. Its strength depends on how much information the remaining features carry about A, which can be estimated by training a model to predict the protected attribute from the other inputs.
| Feature | Attribute it may reveal | Mechanism |
|---|---|---|
| Postal code | Race or ethnic origin | Residential segregation |
| First name | Sex and ethnic origin | Naming conventions |
| Employment gaps | Sex or disability | Caregiving or medical leave patterns |
| Graduation year | Age | Direct arithmetic relation |
| Purchase history | Religion or health status | Dietary or medical product choices |
Unawareness can even make bias harder to detect and correct, because mitigation methods such as reweighting or group-specific thresholds need the attribute to work.
Handling protected attributes in practice
Responsible handling separates decision-time use from audit-time use. In most regulated settings the attribute should not drive individual decisions directly, yet it should be available, under strict access controls, for measuring and mitigating disparities.
Regulation increasingly reflects this split. The EU AI Act permits providers of high-risk systems to process special categories of data where strictly necessary for bias detection and correction, subject to safeguards such as pseudonymization and deletion after use.
| Situation | Recommended approach | Reason |
|---|---|---|
| Attribute collected with consent | Store separately for audits only | Enables metrics without decision use |
| Attribute not collected | Estimate it with proxy inference for aggregate audits | Allows group-level checks with known error |
| Several attributes matter | Test intersectional subgroups | Harms can hide in combinations such as age and sex |
| Small subgroups | Report confidence intervals | Avoids false conclusions from noise |
Limitations and common misconceptions
Protected attributes are social categories with fuzzy and changing boundaries, while models treat them as fixed labels. Self-identification, mixed backgrounds and non-binary gender identities do not fit neatly into the categorical encodings that many datasets use.
- Equal accuracy is not equal fairness: overall accuracy can be identical while error types differ sharply between groups.
- Legal compliance is not the ceiling: an attribute absent from the law may still be ethically relevant in a given context.
- Inferred attributes carry error: proxy estimates of race or sex can misclassify individuals and should be used only in aggregate.
- One attribute at a time misses intersectional harm, which is a recurring finding in facial analysis audits.
The concept also connects to counterfactual fairness, which asks whether a decision would change if the protected attribute of the same person were different, a question that requires an explicit causal model rather than simple group statistics.
The idea in brief
A protected attribute is a personal trait such as age, sex or a disability. It should not count against someone in a decision. Laws name these traits because people were treated unfairly for them in the past.
Computers now help decide who gets a loan or a job interview. Checking fairness means comparing results for different groups. Hiding the trait from the computer is not enough. Clues like a home address or a first name can still reveal it.
A practical scenario
A mid-sized bank in the United States builds a credit scoring model for personal loans. The model never receives applicants’ race or sex, but the bank’s model risk team must still show that it does not produce unlawful disparate impact on protected groups.
Because race is not collected for most loan products, the team estimates it in aggregate with a surname and geography method and compares approval rates. The analysis shows a selection rate ratio of 0.74 for one group, below the common 0.8 benchmark. A feature review finds that a variable based on the applicant’s branch location acts as a strong proxy for ethnic origin.
The team replaces that variable with a direct measure of repayment capacity and retrains the model. The ratio rises to 0.86 with only a small loss in predictive performance, and the full analysis is documented for regulators and internal audit.