Is every sensitive attribute legally protected?

Not necessarily. Sensitivity and legal protection are related but distinct concepts.

Does removing protected attributes make a model fair?

No. Proxies, biased labels, and workflow barriers can remain.

Can protected-attribute data be useful for auditing?

Yes, when its collection and use are appropriate and safeguarded. Without relevant information, some disparities may be difficult to assess.

AI Glossary / Governance, ethics and safety / Intermediate

Protected attribute

Definition

A protected attribute is a personal characteristic associated with protection against discrimination under an applicable law or policy. In AI fairness work, the term often identifies a characteristic used to examine whether a system treats groups differently. The relevant attributes depend on jurisdiction and context; there is no single universal list that automatically governs every AI application.

What it means

The attribute is the characteristic, such as age or disability status. A group defined by a relevant value or condition is a protected class in the corresponding legal context. Technical discussions sometimes use protected attribute and sensitive attribute interchangeably, but sensitive is broader and may include information that raises privacy or fairness concerns without having the same legal status. An organization must therefore state what it means. A dataset column is only a representation of a characteristic: missing values, coarse categories, self-reporting practices, or incorrect coding can distort the analysis.

How it works

An assessment identifies the applicable context, the characteristics that matter, and how their data can be handled appropriately. Analysts can then compare outcomes, errors, or access across relevant groups while accounting for sample sizes and the decision process. Protected-attribute information might be used in a controlled audit without being an input to operational predictions. Conversely, excluding it from the model does not ensure that related information is absent; location, history, or other features can act as proxies. Fairness assessment must examine the actual system and its effects rather than rely on a feature-list declaration.

Three practical examples

A team checks whether an automated application tool's completion rates differ for users with and without a disability, where appropriate data and safeguards are available.
A hiring-tool audit examines outcomes by characteristics relevant to applicable employment protections, keeping audit data separate from ordinary reviewer access.
A service team reviews whether its chosen age categories hide a problem affecting a narrower age range. These examples illustrate assessment questions, not a claim that collecting or using every characteristic is always permitted.

Think of it this way

Think of a protected attribute as a dimension along which a fairness review must ask a careful question. It resembles choosing the right headings for an inspection sheet, but the heading alone does not perform the inspection. The analogy is limited because people can belong to intersecting groups, categories may be socially and legally complex, and reducing a person to one recorded field can miss important context.

Worked example

Suppose a hypothetical service audit records completion for two groups defined by an attribute relevant to its review. Group A has 180 completions among 200 attempts, or 90%. Group B has 70 completions among 100 attempts, or 70%. The observed gap is 20 percentage points. Removing the attribute column from the production model would not change these recorded outcomes or explain their cause. The team should investigate where failures occur, whether the audit categories and records are reliable, and whether missing attempts were excluded. The figures establish an observed disparity in this sample, not a causal finding or a legal determination. If 50 additional attempts lack group information, the report should show that limitation rather than silently assign them to either group.

Illustrative case study

Fictional case: Harbor Forms notices uneven completion of an AI-assisted registration flow. It conducts a carefully governed accessibility study and discovers that one input method is difficult for some participants. The team corrects the interface and repeats the evaluation. This invented case illustrates why fairness concerns can arise before a predictive model makes any decision. Recording an attribute can help reveal a problem, but the response should address the actual barrier rather than assume that changing the model alone is sufficient.

Common misconceptions and mistakes

Protected does not mean that the characteristic may never be recorded for any purpose. Whether collection or use is permitted depends on the law, purpose, and safeguards. Nor does protected mean that every difference in outcomes has already been proven unlawful. Equal aggregate rates also do not establish equal treatment in every case. Broad categories can hide differences at intersections, such as age and disability together. Inferring an unrecorded attribute from other data can introduce additional errors and privacy concerns; such inference should not be treated as equivalent to reliable, appropriately collected information.

Business use

Define the relevant protections with qualified legal and policy input for the intended setting. Document why each attribute is needed, who can access it, and how it will be represented and retained. Distinguish audit use from decision use and test the technical controls enforcing that distinction. Report group sizes, uncertainty, missingness, and meaningful intersections. Investigate upstream access barriers and downstream human actions alongside model outcomes. The U.S. EEOC's employment guidance provides one context-specific example of protected characteristics, including race, religion, sex, national origin, certain age protections, disability, and genetic information; other contexts require their own analysis.

Check your understanding

A developer says a model cannot produce discriminatory outcomes because it never receives a protected-attribute field. Is that enough evidence?

Answer: No. Other inputs can carry related information, and harms can arise from labels, sampling, or the surrounding process. The system's effects still require assessment.

Related terms

Protected class
Sensitive attribute
Group fairness
Disparate impact
Missing data
proxy variable
intersectionality
data minimization
fairness audit

Learn next

Choose a hypothetical decision context and separate three questions: which characteristics have applicable protections, which data would support a useful audit, and which uses of that data are appropriate. Document the gaps before selecting a fairness metric.

Fairness metric

Sources

U.S. EEOC, who is protected from employment discrimination: https://www.eeoc.gov/employers/small-business/3-who-protected-employment-discrimination Google Machine Learning Glossary, sensitive attributes and proxies: https://developers.google.com/machine-learning/glossary Fairlearn, fairness assessment: https://fairlearn.org/main/user_guide/assessment/perform_fairness_assessment.html

Previous: Prospective study Next: Protected class