Fairness in Artificial Intelligence (AI) and Machine Learning (ML) is a recent and hot topic. As ML models are used in insurance pricing, the fairness topic also applies there. Just last month, Lindholm, Richman, Tsanakas and Wüthrich published a discussion paper on this subject that sheds new light on established AI fairness criteria. This post provides a short summary of this discussion paper with a few comments of my own. I recommend the interested reader to jump to the original: A Discussion of Discrimination and Fairness in Insurance Pricing.
First of all, I’d like to state that fairness in the form of solidarity and risk sharing was always at the heart of insurance and, as such, is very very old. The recent discussions regarding fairness has a different focus. It comes with the rise of successful ML models that can easily make use of the information contained in large amounts of data (many feature variables). A statistician might just call that multivariate statistical models. Insurance pricing is a domain where ML models (including GLMs) are successfully applied for quite some time (at least since the 1990s), and where at the same time protected information like gender and ethnicity might be available in the data. This led the European Council to forbid gender in insurance pricing.
The important point is—and here speaks the statistician again—that not using a certain features does in no way guarantee that this protected information is not used by a model. A car model or type, for instance, is correlated with the gender of the owner. This is called proxy discrimination.
The brilliant idea of Lindholm et al. was to construct an example where a protected feature does not influence the actuarial best price. So, everyone would agree that this is a fair model. But it turns out that the most common (statistical) definitions of AI fairness all fail. All of them judge this best price model as unfair. To be explicit, the following three group fairness axioms were analysed:
- Independence axiom / Statistical parity / Demographic parity
- Separation axiom / Equalized odds / Disparate mistreatment
- Sufficiency axiom / Predictive parity
On top of that, these 3 fairness criteria may force different insurance companies to exclude different non-protected variables from their pricing models.
How to conclude? It turns out that fairness is a complicated matter. It has many sociological, cultural and moral aspects. Apart from this broad spectrum, one particular challenge is to give precise mathematical definitions. This topic seems to be, as the paper suggests, open for discussion.
Leave a Reply