Does removing the sensitive attribute remove its information?

Not necessarily. Other inputs can contain correlated information.

Is every proxy relationship perfect?

No. A variable can function as a partial stand-in and still meaningfully affect outcomes.

Must every correlated feature be deleted?

Not automatically. The appropriate response depends on the purpose, actual effects, alternatives, legal context, and the fairness concern being addressed.

AI Glossary / Governance, ethics and safety / Intermediate

Proxy (sensitive attributes)

Definition

A proxy for a sensitive attribute is another variable that serves as a stand-in for that attribute. It may carry information about group membership even when the sensitive attribute itself is absent from a model's inputs. For example, a location variable can be associated with demographic characteristics in a particular population. The strength and implications of that association depend on context.

What it means

A model does not have to receive an explicit group label to produce outcomes related to that group. Other inputs, individually or in combination, can contain similar information. A proxy can be deliberately chosen or can emerge through the relationships in data. The term does not mean that the substitute perfectly identifies every person, nor that every correlated feature is automatically inappropriate. The important questions are what information the feature carries, how the system uses it, and what consequences follow. This entry concerns proxies for sensitive attributes, not network proxies or proxy labels used to approximate prediction targets.

How it works

Statistical relationships connect the proxy variable with the sensitive characteristic. During training, a model may learn to use the proxy because it predicts the target or historical labels. If those labels reflect unequal treatment, the resulting pattern can reproduce or amplify a disparity. Several seemingly weak proxies can jointly reveal substantial information, so checking one feature at a time may miss the issue. Correlation is evidence of association, not proof of a particular causal pathway. Removing a suspected proxy may change outcomes, but it does not establish that all related information or all unfairness has disappeared.

Three practical examples

A service model uses a neighborhood code that is strongly associated with demographic composition in its deployment area.
A language-based system learns patterns associated with a user's background and produces different service recommendations, even without an explicit demographic field.
A fairness audit examines whether a combination of education history and location carries information about a sensitive characteristic. These examples describe potential mechanisms that require evidence; they do not assert that the listed features always act as harmful proxies.

Think of it this way

Think of a proxy as an indirect clue in a puzzle. Removing the label on the answer does not prevent someone from inferring it from the remaining clues. The analogy is limited because statistical clues are uncertain, vary between populations, and can be useful for legitimate purposes. A responsible analysis must examine the decision process and harm rather than treating every clue as conclusive evidence of intent or identity.

Worked example

Suppose a hypothetical dataset contains 100 people in Group A and 100 in Group B. Eighty people in A and twenty in B live in Area X; the remaining people live in Area Y. A simple decision rule selects everyone in X and nobody in Y. It never reads the group field, yet it selects 80% of A and 20% of B, a 60-percentage-point gap. Area is an imperfect stand-in: X includes members of both groups. This arithmetic demonstrates how a group disparity can arise without directly using the group label. It does not determine whether the rule is justified, causal discrimination has occurred, or a law was violated. Those questions require the real purpose, evidence, alternatives, and applicable framework.

Illustrative case study

Fictional case: River Support removes a demographic field from an experimental allocation model and assumes the fairness issue is resolved. An audit finds that location and past service history still produce markedly different recommendations. The team investigates how those records were generated, compares alternatives, and examines whether access barriers influenced the historical labels. This invented case illustrates why proxy analysis must consider data generation and outcomes, not just whether a sensitive column appears in the training file.

Common misconceptions and mistakes

A proxy does not need to predict a sensitive attribute perfectly. Intent is also not required for a statistical proxy relationship to exist. Conversely, a relationship with a sensitive attribute is not by itself proof that a model uses that relationship in a harmful way. Feature importance alone may not reveal the full mechanism, especially with correlated inputs. Deleting one variable can shift reliance to another. It is also risky to treat inferred group membership as certain when conducting an audit; estimation errors can distort measured disparities and affect privacy.

Business use

Begin with the harms and decisions that matter, then investigate plausible proxy pathways. Examine relevant associations, model behavior, data provenance, and group outcomes using appropriate safeguards. Compare alternative features or processes, but assess usefulness and harms together rather than optimizing a single correlation statistic. Where sensitive attributes are unavailable, state the limits of fairness claims instead of assuming unawareness is sufficient. If estimated group membership is used for research or auditing, document its uncertainty and justification. Revisit the analysis when the population or feature collection changes, because proxy relationships may change too.

Check your understanding

A model selects people solely by area, and area is strongly associated with a sensitive group. Can the model create a group disparity without reading the group label?

Answer: Yes. The area variable can act as a proxy, as the worked example shows. The observed mechanism still needs contextual assessment before broader conclusions are drawn.

Related terms

Learn next

For one hypothetical sensitive attribute, list two possible indirect signals and explain why their relationship may vary by population. Design an audit that measures outcomes and uncertainty, then distinguish the evidence of association from any claim about causation or intent.

Sensitive attribute

Sources

Google Machine Learning Glossary, proxy and unawareness entries: https://developers.google.com/machine-learning/glossary Fairlearn, assessing fairness in context: https://fairlearn.org/main/user_guide/assessment/perform_fairness_assessment.html NIST, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence: https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=934464

Previous: Proxy metric Next: Pruning