I was invited to give a talk at the “IFoA AI Ethics Governance & Risk Management forum” at the end of the week. I uploaded some slides to start the discussion.

I will not give a highly technical lecture, but rather to propose a way of thinking about fairness in insurance AI that may be useful for actuaries, for governance people, and for anyone involved in model risk or oversight. The main message is simple. In insurance, fairness cannot be reduced to predictive accuracy. And once we move to modern AI and large data environments, simply removing a sensitive variable is no longer enough to prevent discrimination. So the real question becomes: how should we govern the trade-offs between fairness and accuracy, and different notions of fairness?

Let me start with a very simple example, illustrating the fact that simply removing a sensitive variable is not enough to prevent discrimination. Here I take a French motor insurance dataset. In the raw data, average claim frequencies are about 8.94% for men and 8.20% for women. Now suppose that I estimate a logistic model for annual claim frequency, but I explicitly exclude gender from the explanatory variables. At first sight, this looks like the fair thing to do. I am not using gender, so maybe I am safe. But what happens?
As I add more and more variables to the model, the predicted frequencies for men and women move closer and closer to the empirical frequencies in the data. With enough explanatory variables, I am basically reconstructing the original gap. So the model is not using gender explicitly, but it is learning from variables that are correlated with gender. In other words, the information has not disappeared. It has simply been redistributed across other variables.
That is why I think the old defence of ‘we do not use the sensitive variable, therefore the model is fair’ has become much weaker in a big-data context (and here, it’s not big, it’s only 15 basic explanatory variables). Modern predictive models are very good at finding statistical associations. If protected traits leave traces in the data, the model will usually pick them up.
And this is not only a technical issue. It is also a philosophical one. The model tends to reproduce what is in the data, that’s what we call ‘generalization’ in machine learning. If historical inequalities are present in the data, the model will often learn them very efficiently. This is why I mention Hume here, and the old ‘is-ought’ problem. From the fact that a disparity exists in the data, it does not follow that it should be reproduced in pricing. So this slide is really the starting point for the rest of the talk: fairness through unawareness is often not fairness at all; it is frequently a way of hiding the channel through which unfairness reappears.
So once we accept that unawareness is not enough, we need to ask a more difficult question: what exactly do we mean by fairness?

Before going to the recent actuarial papers, I want to spend one minute on a classical dilemma. This is not new. In legal and philosophical debates, there has long been a tension between two positions.
One position says: if you want to correct unequal outcomes, you may need to take account of the protected characteristic. That is the logic behind many discussions of affirmative action or corrective treatment. If the world is already unequal, pretending not to see the relevant characteristic may simply reproduce the inequality. The opposite position is the so-called colorblind view: the way to stop discrimination is to stop using the sensitive characteristic altogether. That sounds very appealing, because it seems neutral and simple.
The trouble is that in insurance, and especially in AI-based pricing, both positions run into difficulties. If you ignore the sensitive variable, you may still reproduce its effects through proxies, that’s what we’ve seen before. But if you use it for corrective purposes, you are entering a difficult space involving direct differential treatment, redistribution, and legal or ethical contestability.
So what I want to emphasize here is that the fairness debate is not just about technical metrics. It is rooted in a much older normative tension: should fairness mean treating everyone under the same formal rule, or should it sometimes mean treating people differently in order to offset structural inequalities? I think that this dilemma is exactly why actuarial fairness debates cannot be settled by model performance alone.
In the contexte of insurance, the recent SOA report discussed in a previous meeting stresses this important tension fairness is not one thing.

The report is useful because it is very practical and very clear. Its main point is that fairness in life insurance is not a single concept. There is no universal fairness metric that would solve the problem once and for all. Instead, the paper distinguishes between different families of fairness criteria.
A first distinction is between individual fairness and group fairness. Individual fairness is close to the actuarial intuition: similar risks should be treated similarly. Group fairness, by contrast, is about parity across groups. Those two perspectives can both sound reasonable, but they are often in tension. The paper also reviews several criteria. On the individual side, it discusses unawareness, awareness, and omitted-variable bias. On the group side, it discusses independence, sufficiency, and separation.
The most useful takeaway is not any particular metric. It is the methodological lesson: before choosing a metric, you must define what notion of fairness you are trying to serve. Otherwise you are optimizing a number without knowing what ethical or institutional objective it really represents. The SOA paper gives us a map. It tells us: do not look for the one true fairness metric. Clarify the objective first, and then be explicit about the trade-offs.”
Our own recent work tries to push that argument one step further, by asking what the core dimensions of fairness in insurance pricing really are…

In our recent paper on what we call, with Olivier Côté, and Marie-Pier Côté, “the fairness trilemma“, we recall the idea that insurance is not just a predictive exercise. It is also a social institution. Insurance has a double nature. On the one hand, it is about risk-based pricing: aligning premiums with expected losses. On the other hand, it is also about risk-sharing: spreading burdens, preserving access, and sometimes accepting cross-subsidies. From there, we argue that fairness in insurance pricing is governed by a trilemma between three principles.
- The first is actuarial fairness. In simple terms, this means that premiums should reflect expected losses as accurately as possible.
- The second is social solidarity. This means that we may accept departures from strict risk-based pricing in order to preserve access, affordability, or some broader social objective.
- The third is causal legitimacy. This is extremely important in the AI context. Not every predictive variable is equally legitimate. A variable may improve predictive performance, but if it works mainly by reconstructing a protected characteristic, or if its causal meaning is questionable, then its legitimacy is fragile. So causal legitimacy asks not only whether a factor predicts, but whether it deserves to shape the price.
The key point is that these three principles are all attractive, but they cannot be fully satisfied at the same time. If I strengthen actuarial fairness, I may weaken solidarity. If I enforce more solidarity, I move away from pure risk adequacy. If I become very demanding about causal legitimacy, I may sacrifice some predictive performance and perhaps some segmentation logic. So the problem is not to find the perfectly fair model. The problem is to govern the trade-offs. That is why I insist on governance tools: scorecards, causal due diligence, comply-or-explain procedures. The point is to make choices explicit, reviewable, and contestable. In that sense, fairness is not just a modelling problem. It is a governance problem.
But of course, at this stage, someone in the room may say: all this sounds interesting, but how do we make it operational? How do we move from philosophy to something an actuary can actually measure?

That is exactly the purpose of the CAS paper. The starting point is again that removing a sensitive variable does not remove discrimination, because allowed variables may still act as proxies. So the challenge is to expose indirect discrimination in a way that is usable in actuarial practice. The objective of the paper is to make fairness operational and measurable in actuarial terms. It is not enough to speak in abstract statistical language. In insurance, we need to know what unfairness means in premiums, in dollars, and across segments of policyholders.
The paper is based on a real auto insurance case study from Québec, using credit score as the sensitive variable. And it again organizes the discussion around three dimensions: actuarial fairness, social solidarity, and causality legitimacy. This is important because it shows continuity with our more conceptual framework. The toolbox is not something separate. It is a way of translating those dimensions into measurable diagnostics. If the trilemma paper says fairness must be governed, the CAS paper says: here are some tools that help you see what you are governing.

One of the most interesting contributions of the toolbox is that it introduces local metrics such as risk spread, proxy vulnerability, fairness range, and parity cost. Let us focus on two of them.
The first is proxy vulnerability. Intuitively, proxy vulnerability measures how much a segment may be over-priced or under-priced because apparently neutral variables indirectly reconstruct the sensitive attribute. This is very valuable because it moves us away from vague concerns about bias and toward a concrete actuarial question: where, and by how much, might the pricing rule be unfair?
The second is parity cost. This is also very important because it forces us to be honest. If we want more parity across groups, that is not free. Someone has to bear the cost. In plain English, enforcing one notion of fairness may end up robbing Peter to pay Paul. And I do not say this as a criticism; I say it because redistribution should be explicit rather than hidden.
Another strength of the toolbox is that it helps actuaries identify local pockets of unfairness. A model may look acceptable on average across broad groups, but still be very problematic in specific vulnerable niches. So fairness should not be assessed only globally. Local diagnostics matter.
This is also, I think, very relevant for AI governance. Senior management and boards do not just need a statement saying ‘the model passed a fairness test’. They need to know where the model is fragile, which subpopulations are exposed, what the monetary magnitude is, and what the trade-offs are if a correction is introduced. So in practice, the toolbox helps turn fairness into something discussable in risk committees and governance processes.

To wrap up, first, the broader background for all this is developed in my recent textbook, which tries to connect legal, philosophical, statistical, and actuarial perspectives on discrimination in insurance. Second, fairness should not be assessed only at the level of one insurer’s portfolio. A pricing rule may look fair on one portfolio and still be unfair at the market level, because portfolios are not representative of the insured population. This is a selection-bias problem. If different insurers attract different segments of the market, portfolio-specific fairness does not necessarily aggregate into market-wide fairness. So we also need to ask: fair for whom, and on which reference population?
Third, there is always a cost and a redistribution issue. If we enforce one notion of fairness, someone pays. That does not mean fairness is undesirable. It means that fairness choices are unavoidably political, institutional, and governance choices, not just technical ones. And fourth, there is a major practical difficulty when sensitive variables are unobserved. Very often, people say: let us not collect the sensitive variable. But then how do we audit fairness? How do we provide evidence? How do we challenge the model? If the protected attribute is precisely the thing we do not observe, fairness becomes not just a modelling issue, but a data, audit, and governance issue.
So my overall conclusion would be the following. AI did not create the fairness problem in insurance, but it has made it sharper. It has weakened the old comfort of unawareness. It has increased the power of proxies. And it has made the trade-offs more difficult to hide. For actuaries, I think this creates both a risk and an opportunity. The risk is to treat fairness as a box-ticking exercise or as a purely technical constraint. The opportunity is to contribute something distinctive: a language of prices, trade-offs, portfolios, and governance. So perhaps the right ambition is not to promise a perfectly fair model. It is to build institutions that are capable of identifying proxy discrimination, making normative choices explicit, and governing them responsibly.