Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When AI Predicts Your Life Expectancy for Insurance
Transcript
- Lucas: Luna, let me ask you something. If a machine learning model looked at your social media posts and your genetic data to predict how long you'll live — and then your insurance premium went up because of it — does that feel like innovation or surveillance? Luna: Honestly, it feels like a dystopian pitch. But I know this is already happening. Where do we start? Lucas: Let's start with a specific case from late 2025. A major European insurer — let's call it InsureCo — rolled out a new life insurance product. The pricing was determined by an AI that ingested medical records, yes, but also fitness tracker data, grocery purchase history, and even public Instagram posts. The model's explicit goal was to predict 'longevity risk'. Luna: That's a lot of data. Did customers opt into all of that? Lucas: That's the first flashpoint. The terms of service were buried in a 45-page document. Most people just clicked 'agree'. But here's where it gets concrete: a 34-year-old cyclist named Marie from Lyon applied for a policy. She's a marathon runner, healthy, non-smoker. Her Instagram is full of race photos. The model flagged her as 'high risk — elevated mortality' because her posts showed participation in endurance sports, which the algorithm correlated with 'risk-seeking behavior' and possible cardiac stress. Luna: Wait — so being fit and active was treated as a liability? That's backwards. Did she appeal? Lucas: She did. And under the EU's AI Act, which came into full enforcement in phases through 2026, she had a 'right to explanation'. The insurer had to give her a meaningful account of how the model reached its decision. What they gave her was a printout listing seven factors, including 'social media activity related to physical exertion' and 'purchases of high-protein foods' — which they apparently linked to a profile of 'compensatory health behaviors'. Luna: So the algorithm basically penalized her for living a healthy life, just framed as 'statistical correlation'. Did the regulator step in? Lucas: The French data protection authority, the CNIL, opened an investigation. They found that the model had a 12 percent error rate in predicting five-year mortality for women under 40 — but the premium adjustments were applied as if the predictions were near-certain. The insurer settled and agreed to pause the product. But the underlying question remains: can AI ever be actuarially fair when it's trained on biased proxies? Luna: And that's the core tension. Traditional underwriting uses age, gender, smoking status — coarse but statistically defensible. AI promises more precision, but it also brings in data that isn't just correlated with risk, but with lifestyle, race, or socioeconomic status. Lucas: Exactly. One study from the European Insurance and Occupational Pensions Authority found that models using social media data can indirectly infer mental health conditions, political affiliation, and even sexual orientation — all of which are protected under EU non-discrimination law. The insurers argued they weren't using those categories directly, but the model's outputs were clearly correlated. Luna: So the algorithm is basically rediscovering redlining, but with better PR. What's the fix? Lucas: Some regulators are pushing for 'algorithmic impact assessments' before deployment, similar to what Canada proposed in their AI and Data Act. Others want a ban on certain data types entirely — no genetic data, no social media scraping for insurance purposes. The industry argues that would stifle innovation and make pricing less accurate. Luna: Right, but 'accurate' doesn't mean 'fair'. If the model is accurate but systematically disadvantages people who post about their hobbies online, that's a choice about what kind of risk pool you want. Lucas: And that's the ethical knot. Insurers have a fiduciary duty to price risk accurately — otherwise, low-risk policyholders subsidize high-risk ones. But when accuracy comes from scraping your Instagram, you're crossing into a different kind of territory. Luna: Speaking of territory — I want to pivot slightly, because this connects to something I've been thinking about. A lot of these ethical conversations are only possible because the show is ad-free and listener-supported. We don't have a sponsor telling us to soft-pedal the critique. Lucas: That's real. Look, a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that is literally what funds making this many episodes. It keeps us independent. Luna: Exactly. So if today's conversation gave you something usable, that's the reason. No ads, no pressure — just people who care about these questions. Lucas: Right. And on that note, back to Marie's case. The CNIL eventually ruled that the insurer had to delete the social media data and recalculate her premium using only traditional actuarial factors. Her premium dropped by 40 percent — back to what it should have been. Luna: That's a win for transparency. But it's one case. How many people never find out they've been profiled? Lucas: That's the systemic problem. The AI Act requires a 'right to explanation' only for decisions that are 'solely automated' and produce 'legal effects or similarly significant effects'. But insurers are already building hybrid models where a human rubber-stamps the AI output — which might exempt them from the full transparency requirements. Luna: So the human-in-the-loop can become a loophole. If the human just clicks 'approve' on 99 percent of recommendations, is that really human oversight? Lucas: Exactly. The European Commission's own guidance says oversight must be 'meaningful' — but enforcement is patchy. Meanwhile, in the US, there's no equivalent federal law. The NAIC — National Association of Insurance Commissioners — issued model bulletins in 2024, but they're voluntary. Luna: Let's talk about an alternative. Is there a way to use AI for underwriting ethically? Lucas: Some startups are trying 'fairness-constrained' models — they train the algorithm to minimize predictive disparity across demographic groups, even at the cost of overall accuracy. One uk based company, FairRisk, published a paper showing they could reduce racial bias in auto insurance pricing by 60 percent while only increasing aggregate error by 3 percent. Luna: That's promising. But does that actually get adopted, or does it stay in academic papers? Lucas: Adoption is slow. Incumbents argue that any accuracy loss hurts their competitiveness. But the real barrier might be legal. If a fairness-constrained model charges a smoker less because it's compensating for other factors, is that 'unfair discrimination' against non-smokers? The definitions get tangled. Luna: So 'fairness' itself is a contested term. For regulators, it often means treating like cases alike. For ethicists, it might mean ensuring outcomes don't disadvantage historically marginalized groups. Those definitions can conflict. Lucas: And that's why Marie's case is such a useful lens. It's not an edge case — it's the central tension. The insurer genuinely believed they were being more precise. But precision without fairness is just a better map of an unjust territory. Luna: I want to push on something else. The AI model in this case used grocery purchase data. What does that even capture? Lucas: The insurer partnered with a loyalty-card data aggregator. Purchases of organic vegetables, for example, were correlated with lower mortality — so those customers got a discount. But purchases of frozen meals or soda were correlated with higher mortality. The problem: grocery patterns are heavily influenced by income and geography. In a food desert, you may not have access to fresh produce. So the model effectively penalizes low-income neighborhoods. Luna: That's a classic bias cascade. The data reflects structural inequality, the model learns it, and then the pricing reinforces it. Lucas: Bingo. And this is where the EU AI Act's 'high-risk' classification matters. Insurance underwriting is considered high-risk, which means providers must conduct a fundamental rights impact assessment before deployment. But the French case revealed that InsureCo's assessment was essentially a checkbox exercise — they hadn't tested for proxy discrimination. Luna: So the regulation is there on paper, but the enforcement infrastructure isn't fully mature. How do we fix that? Lucas: Some advocates are calling for mandatory 'algorithmic audits' by independent third parties before any model goes live, similar to financial audits. The EU is piloting a voluntary certification scheme, but it's not mandatory until at least 2027. In the meantime, whistleblowers and journalists have been key to surfacing problems. Luna: And that's where the podcast ecosystem has a role — we can highlight specific cases so listeners know what to look for. If you get a surprise premium increase, ask for the explanation. Exercise that right. Lucas: Exactly. And if the explanation is vague or algorithmic, file a complaint with your data protection authority. The GDPR already gives you tools — the AI Act layers on more. Luna: Before we wrap, I want to circle back to something you said earlier about the cyclist Marie. The model essentially pathologized her healthy behavior. That's a reminder that AI doesn't just reflect biases — it can create new ones by redefining what 'normal' looks like. Lucas: That's a great way to put it. The actuarial profession has centuries of experience defining risk pools. AI is rewriting those definitions at machine speed, without the professional norms. And that's both the promise and the danger. Luna: So the bottom line: AI in insurance can improve accuracy, but only if we constrain the data inputs, mandate transparency, and enforce fairness metrics. Otherwise, it's just surveillance with a premium attached. Lucas: And the next time you hear an insurer say 'we use AI to offer you a personalized price,' ask yourself: personalized for me, or personalized against me?