Van Bekkum, a lawyer and computer scientist, investigated how insurers use artificial intelligence and how customers react to it. To this end, he not only examined existing cases and current regulations, but also conducted a survey among Dutch people. The results show that many people find the data-driven price differences in insurance policies incomprehensible and often unfair.
Price difference based on house number
‘It can sometimes be difficult to fathom,’ says Van Bekkum, who cites as an example a case uncovered several years ago by the Consumers’ Association. ‘When taking out car insurance, some insurers look at your house number. That house number largely determines how much you pay per month. According to insurers, there is a statistical correlation behind this, for example with income or crime rates in a neighbourhood. The algorithm might determine that someone with a very high house number lives in a block of flats where criminals are more likely to reside, so the price is then increased. That correlation may well exist, but the public doesn’t get that story: they only see the price difference. It is precisely that lack of transparency that makes it all problematic.’
A major problem is that many of these differences do not fall under existing anti-discrimination legislation. Van Bekkum therefore distinguishes between discrimination and unfair differentiation. ‘Discrimination occurs when it is based on a characteristic protected by law, such as ethnicity. But if you make a distinction based on someone’s house number, that is not a protected characteristic. Whilst the underlying algorithm may well identify all sorts of correlations that result in vulnerable groups having to pay more for their insurance on a structural basis.
Politicians tackling the symptoms
Although this is therefore incredibly unpopular with the Dutch public, the government is doing little about it for the time being. Van Bekkum: ‘Non-life insurance is part of a free market, so the government lets companies do as they please. The House of Representatives has looked into protecting certain characteristics such as income, but that is merely treating the symptoms: you can ban a few characteristics, but the algorithm will always find other characteristics on which it can unfairly differentiate.’
Due to the popularity of ChatGPT and other chatbots, interest in AI has risen significantly in recent years. Insurers are also enthusiastic and are looking with interest at algorithms based on large language models (LLMs), says Van Bekkum. ‘But those LLMs actually only increase the risks. They are models that merely predict the next word in a sentence based on existing text; they understand nothing about how correlations arise or why a particular prediction is made.’
If algorithms increasingly determine who pays how much, without this being properly explainable to customers or even to insurers themselves, confidence in insurance is under pressure, warns Van Bekkum. ‘We must prevent ending up in a situation where differences keep growing, whilst nobody can explain how they came about.’