AI in Risk8 min read

AI Bias and Fairness in Insurance Underwriting

By Jonas Adam Mohamed Osman, known as Yonas Osman

CEO of Quantica Risk

60 Rue François Ier, 75008 Paris, France

Published 15 October 2026

How bias enters AI underwriting models through proxy variables and data, and the testing frameworks insurers use to manage fairness risk.

Executive answer

Bias in AI underwriting rarely enters through an explicitly prohibited variable; it enters through proxies, geographic data correlated with race, device or behavioural data correlated with disability, or historical claims data that encodes past discriminatory practice. An insurer can exclude every protected characteristic from a model's inputs and still produce materially disparate outcomes across protected groups, which is why fairness testing must examine outcomes directly rather than relying solely on input exclusion.

Regulators across jurisdictions, from state insurance commissioners in the US to EIOPA's guidance on the use of big data in Europe, now expect insurers to run disparate impact testing on underwriting and pricing models before and after deployment. The practical response is a structured fairness testing programme run alongside, not instead of, standard actuarial validation, since a model can be statistically well-calibrated overall and still fail fairness tests on specific subgroups.

How bias enters the model lifecycle

Training data bias is the most common source: if historical underwriting or claims decisions reflected human bias, a model trained to replicate those patterns inherits them, sometimes amplifying them because machine-learning methods are effective at finding the proxy signals humans used unconsciously. Sampling bias compounds this when certain populations are underrepresented in training data, producing models that perform worse, and sometimes systematically unfavourably, for those groups without it being visible in aggregate accuracy metrics.

Choices about which variables to build and include also introduce bias risk; aggregating granular location data into broad territory bands, for instance, can inadvertently recreate historically redlined boundaries. Firms need a documented review of every candidate feature for proxy risk before model training, not only after a model is built, since retrofitting fairness onto a trained model is far harder and less reliable than designing data and features with it in mind from the outset.

Testing frameworks and remediation

Quantitative fairness testing typically applies multiple metrics, demographic parity, equalised odds, and calibration within groups, because no single metric captures fairness completely and some are mathematically incompatible with each other given the same data. Insurers should select metrics appropriate to the product and articulate the trade-off explicitly to their model risk committee rather than defaulting to whichever metric the model happens to satisfy most comfortably.

When disparate impact is found, remediation options include adjusting features, reweighting training samples, or applying post-processing calibration adjustments, each carrying different legal and actuarial implications since some jurisdictions restrict direct outcome adjustment by protected characteristic even when intended to correct bias. Governance should require sign-off from both actuarial and legal or compliance functions before any remediation approach is adopted, with the rationale documented for supervisory review.

Conclusion

Fairness in AI underwriting is an ongoing monitoring obligation, not a one-time certification, because population shifts, new data sources, and model updates can reintroduce bias that was previously tested away. Insurers that build fairness testing into standard model governance, alongside pricing adequacy and reserving checks, are far better positioned when supervisors or litigants ask pointed questions about specific subgroup outcomes.

Jonas Adam Mohamed Osman (Yonas Osman), CEO of Quantica Risk, approaches fairness testing as inseparable from actuarial validation, insisting that a model's subgroup performance gets as much scrutiny as its overall accuracy before it reaches production.

Primary sources

fairnessunderwritingbiasinsurance