A variable that stands in for a sensitive or regulated attribute without naming it directly. Proxy features can carry historical bias into model predictions, which is why fairness review must inspect feature influence, not just the presence or absence of protected labels.
Expanded Definition
A proxy feature is any variable that appears innocuous on its face but closely tracks a sensitive attribute, regulated class, or other hidden factor that should not drive a decision. In AI governance, the key issue is not whether a protected label is explicitly present, but whether another feature behaves as a surrogate and shapes model output in the same direction. That is why fairness review must look at correlation, feature importance, and downstream effect together, rather than treating label removal as sufficient mitigation. Guidance in the NIST AI Risk Management Framework and related fairness practices treats this as a model risk issue, because proxies can preserve historical bias even after protected fields are excluded. See also the broader governance approach in the NIST Cybersecurity Framework 2.0 for accountability and risk management expectations.
Definitions vary across vendors and audit teams on where the line sits between a useful predictive signal and an impermissible proxy, so the term is best handled as a risk condition rather than a fixed category. A postal code may be legitimate for logistics, but in a lending model it may also encode race, income, or neighborhood segregation. The most common misapplication is assuming fairness is solved by deleting protected attributes, which occurs when correlated features remain in the training data and continue to reproduce the same outcome.
Examples and Use Cases
Implementing proxy feature review rigorously often introduces analysis overhead, requiring organisations to weigh model performance and explainability against the cost of additional testing and feature governance.
- A credit model uses postal code as a feature, but the variable acts as a stand-in for race and neighborhood wealth.
- An employment screening model uses university name or graduation year, which may proxy for socioeconomic background or age.
- A fraud model uses device type or time zone, which can correlate with geography, travel patterns, or customer segment membership.
- A healthcare model uses insurance product or clinic location, which can indirectly encode access disparities and protected status.
- A recommendation system uses browsing history in a way that amplifies prior exposure patterns, creating a proxy for demographic or interest-based segmentation.
For AI governance teams, proxy analysis is often paired with feature attribution, counterfactual testing, and slice-based performance checks. The challenge is not always to eliminate every surrogate variable, but to determine whether its influence is justified, documented, and proportionate to the decision context. In many organisations, the question becomes whether the feature is operationally necessary or merely convenient. Where model governance is still maturing, a practical starting point is to catalogue high-risk features and review them against the accountability expectations reflected in the NIST Cybersecurity Framework 2.0 and adjacent AI oversight practices.
Why It Matters for Security Teams
Proxy features matter because they can create hidden discrimination, compliance exposure, and reputational harm even when protected data is never explicitly used. For security and governance teams, this is a control problem as much as a model problem: if feature lineage, training data provenance, and review evidence are weak, it becomes difficult to show that automated decisions were designed and tested responsibly. That is especially important where AI systems interact with identity verification, access decisions, fraud scoring, or agentic workflows, because a proxy can shape who is challenged, who is denied, and who receives elevated trust. In those settings, proxy risk becomes part of broader model governance, not a niche fairness concern. The issue also connects to data minimisation and purpose limitation, since unnecessary features expand both privacy and bias exposure.
Practitioner insight: organisations typically encounter the operational impact of proxy features only after a complaint, audit finding, or adverse decision review, at which point the model can no longer be defended by saying the protected label was absent.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST AI RMF, NIST CSF 2.0, NIST SP 800-63 and NIST AI 600-1 set the technical controls, while EU AI Act define the regulatory obligations.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST AI RMF | AI RMF addresses bias, validity, and accountability issues that proxy features can create. | |
| NIST CSF 2.0 | GV.RM | CSF 2.0 risk management governance fits proxy feature review and accountability. |
| NIST SP 800-63 | Digital identity decisions can be distorted when proxy features influence verification or assurance. | |
| NIST AI 600-1 | GenAI guidance covers harmful outputs and bias risks that can emerge through proxy features. | |
| EU AI Act | The EU AI Act requires risk controls for systems where proxy bias may affect decisions. |
Review proxy features in model governance, document their impact, and test for unfair outcomes before deployment.
Related resources from NHI Mgmt Group
- When is a reverse proxy better than a VPN for access control?
- When does browser automation become a governance problem instead of a productivity feature?
- What is the difference between a managed gateway and a reverse proxy in front of a gateway?
- How can teams spot proxy abuse on compromised Linux systems?
