Removing the protected variable from the model does not make the model fair. That single technical fact has done more to reshape insurance fairness work in the last three years than any regulatory pronouncement, because it forces a harder question — when does legitimate risk classification turn into unfair discrimination, and who decides?

By Elana Hörstmann, Quantitative Specialist

Insurance has always involved segmentation. The actuarial principle is that premiums should reflect expected cost. But modern data science makes segmentation far more powerful, and modern models can infer protected attributes indirectly through proxies. The result is that fairness has stopped being a peripheral ethics topic. It is now part of actuarial model quality.

Three definitions, and why all three matter

Direct discrimination uses a protected characteristic in a decision in a way that is prohibited or unjustified. Indirect discrimination uses an apparently neutral rule with a disproportionate adverse effect on a protected group. Proxy discrimination uses a non-protected variable, or combination of variables, that effectively reconstructs a protected characteristic. Geography, occupation, device behaviour, transaction patterns or distribution channel may carry information about race, income, disability or gender. Whether such use is acceptable depends on context, law, proportionality, justification and governance.

The practical difficulty is that insurance pricing frequently depends on variables correlated with socio-economic conditions. A model can be technically accurate and still create fairness concerns. Lindholm and co-authors give rigorous definitions and a discrimination-free pricing formula that can work across statistical models. Subsequent work shows that simply removing protected attributes is insufficient because other variables can proxy for them. The Actuary Magazine’s framing is the one to internalise — fairness in pricing is a multidisciplinary effort spanning actuarial science, computer science, law and economics. It cannot be delegated only to the modellers or only to the lawyers.

The South African overlay

South African insurers operate under several overlapping frameworks. POPIA section 71 restricts decisions based solely on automated processing where those decisions have legal or similarly significant effects, subject to exceptions and safeguards — relevant where AI materially influences underwriting, pricing, claims repudiation or benefit eligibility. The FSCA / Prudential Authority’s 2025 AI report highlights explainability, disclosure, fairness, operational risk and the ability to challenge decisions. Treating Customers Fairly applies on top of all of that.

The practical implication: insurers should not wait for a model to be challenged before documenting fairness. The fairness policy should be defined before deployment. Where AI is used in the analysis, we apply the controls on our How we use AI page.

A six-step framework that survives a regulator’s reading

1. Define the decision. Pricing, underwriting, claims triage, fraud detection and retention modelling have different legal and ethical exposures. The framework cannot be one-size-fits-all.

2. Identify protected or sensitive characteristics relevant to the jurisdiction and product. In South Africa that means constitutional equality principles, sector regulation, POPIA, insurance conduct rules and product-specific risks.

3. Map proxies. Identify variables that strongly correlate with protected characteristics. The variables do not automatically have to be removed — but their use has to be justified and monitored.

4. Test outcomes. Group-level impact, error-rate differences, calibration by group, adverse impact ratios, sensitivity-based discrimination measures. Where protected-class labels are unavailable, careful proxy or imputation methods may be needed — with legal and ethical oversight, not as a modelling shortcut.

5. Mitigate where appropriate. Variable restrictions, monotonic constraints, post-processing, model simplification, caps, underwriting rules, additional human review, or alternative product design.

6. Document the judgement. Fairness often involves trade-offs between metrics that cannot all be satisfied simultaneously. A committee should record why a variable is appropriate, what alternatives were considered and how customer outcomes will be monitored.

Why purely technical fairness is not enough

Fairness metrics conflict with each other. A model that satisfies one criterion will fail another. A purely statistical definition may not align with local law, customer expectations or the insurer’s purpose. This is exactly why actuarial governance matters: the actuary can explain risk, uncertainty, model behaviour and pricing consequences; legal and compliance can assess statutory and conduct risk; product and distribution can assess customer impact; senior management can decide risk appetite. None of those is optional in the conversation.

Three worked examples

A pricing committee questions a geospatial rating factor. The answer should cover predictive value, causal plausibility, correlation with protected or socio-economic indicators, rate impact, alternatives considered, underwriting rationale and customer-outcome assessment. Not one of those alone.

A claims team uses AI to flag potential fraud. The governance question includes false-positive rates by customer segment, escalation rules, human review, complaint data, and whether the model unfairly delays legitimate claims.

A retention model identifies likely lapsers and offers selective discounts. The fairness question: are vulnerable customers disadvantaged? Does product value remain appropriate? Are distribution practices consistent with conduct obligations?

What good looks like

The insurers that get this right do not bolt fairness testing onto the end of the modelling process. They build fairness into data design, feature governance, modelling, validation, reporting and monitoring. That is the standard modern AI-enabled insurance requires — and it is exactly the kind of work the actuarial function should be leading, because it sits at the intersection of risk classification and customer outcome that actuaries already understand.

If you want help defining a fairness policy you can stand behind, our Modelling and Validation practice covers the framework, the testing and the documentation in one engagement.

Sources