Artificial intelligence is now part of the everyday insurance toolkit: pricing, fraud detection, automated underwriting and claims management all rely on increasingly powerful predictive models. With that power, however, comes a familiar tension: the most accurate models also tend to be the most opaque. Responsible artificial intelligence is the discipline that seeks to reconcile performance with transparency, fairness and accountability.
The black box problem
A linear or logistic regression model is interpretable by construction: each coefficient has a direct meaning. A gradient boosting model with thousands of trees or a deep neural network, by contrast, works as a black box: it predicts well, but it does not readily reveal why it assigned a high premium or declined an application.
In insurance, this opacity is not merely a technical inconvenience. Policyholders are entitled to an explanation, regulators require insurers to show that rates are technically sound and not unfairly discriminatory, and the insurer itself needs to understand its model in order to manage it. A model that nobody can explain is a model that nobody can govern.
Explainability techniques
In response to this challenge, the field of explainable AI (XAI) has emerged, with tools that open the black box without giving up its predictive power:
- SHAP (Shapley Additive Explanations): based on Shapley values from cooperative game theory, it allocates the prediction among the input variables in a consistent way, showing how much each one pushes it up or down. It is now the de facto standard for local and global explanations.
- LIME (Local Interpretable Model-agnostic Explanations): approximates the model's behavior around an individual prediction with a simple, easy-to-read local linear model.
- Partial dependence plots and their individual variants (ICE): show how the average prediction changes as one variable moves while the others are held fixed, revealing nonlinear effects and thresholds.
These techniques are complementary. Partial dependence offers a global view of a variable's effect; SHAP and LIME explain individual cases, which is especially valuable when a specific decision has to be justified to a customer or a supervisor.
Proxy discrimination
A particularly subtle risk is indirect discrimination. Even when a model excludes protected variables (gender, ethnic origin, health status), it can reconstruct their effect through correlated proxy variables: a postal code may correlate with socioeconomic status or origin; certain spending habits may correlate with health conditions. In optimizing its predictions, the model learns the prohibited relationship through the back door.
Detecting these proxies requires deliberate analysis. Leaving out the sensitive variable is not enough: one must measure whether the model's output is capturing it implicitly, and decide, with actuarial and compliance judgment, which correlations are legitimate (tied to real risk) and which amount to unacceptable discrimination.
Fairness as an explicit objective
Algorithmic fairness is an active field that has proposed many formal definitions, including demographic parity, equal opportunity and calibration within groups. An uncomfortable theoretical result is that, in general, these definitions cannot all be satisfied at once: improving one fairness metric usually degrades another. There is no single, universal notion of fairness that a model can simply optimize.
This has a profound implication: fairness is not a purely technical problem but a policy decision that combines ethical, legal and business criteria. Insurance rests on risk differentiation, so the challenge is not to eliminate every distinction but to ensure that distinctions are actuarially justified and socially acceptable.
Model risk and governance
All of the above comes together in the management of model risk: the possibility of losses from decisions based on models that are wrong or misused. Sound governance covers:
- Complete documentation of data, assumptions, limitations and validations.
- Independent validation by a team other than the one that developed the model.
- Ongoing monitoring of performance and data drift in production.
- Traceability and version control, so that the reasons behind each decision can be reconstructed.
Professional frameworks such as actuarial standards on modeling, and growing international regulation of artificial intelligence, point in the same direction: high-impact models require human oversight, transparency and clearly assigned accountability.
The actuary as model owner
This is where the actuarial profession finds a natural and strategic role. The actuary combines a deep understanding of risk, statistical training, regulatory awareness and a code of conduct that demands objectivity and care for the public interest. Bodies such as the Society of Actuaries (SOA), the Casualty Actuarial Society (CAS) and the Institute and Faculty of Actuaries (IFoA) have produced material on data science, fairness and model governance, reinforcing this calling.
The model owner is not a technician optimizing a metric but a professional who answers for the model's consequences: its accuracy, its fairness, its explainability and its compliance. In the age of the black box, responsible artificial intelligence is not about slowing innovation but about governing it rigorously. And in that task, the actuary is called to play a central role.