Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When Your AI Car Insurance Punishes Your Neighbors Driving
Transcript
- Lucas: So I got a message from a listener this week—Sharon, in Kansas City—and she sent me her latest auto insurance renewal. Her premium jumped twenty-two percent. Clean driving record, no claims, no tickets. The reason the insurer gave? 'Increased claim frequency in your geographic area.' Luna: Wait—so she's being penalized for something that happened down the street, not because of anything she did? Lucas: Exactly. A car crash two blocks away, and the AI model that priced her policy said, essentially, 'risk in this zip code is rising, so your rate rises too.' This is not a new practice, but it's becoming way more granular and way more automated. Insurers have always used territory as a rating factor, but historically that meant broad regions—an entire county, sometimes a whole state. Now, with AI and massive location data sets, they're drilling down to the census block group level. Luna: And the AI can do that in real time, pulling accident reports, police data, even traffic camera feeds, and adjusting premiums on the fly? Lucas: That's the direction we're heading. Today, most of the big players—think State Farm, Progressive, Geico—they all use some form of geographic scoring model. The difference is that the newer models incorporate machine learning to identify micro-patterns: a particular intersection that's had three fender benders in a month, or a stretch of road where deer crossings spike in October. And then they feed that into your individual rate calculation. Luna: But if I've never had an accident, why should my premium go up because a neighbor two blocks away is a bad driver? Lucas: The actuarial logic is that you share the same environment—same road conditions, same traffic patterns, same weather exposure, even the same crime risk for car theft. Historically, that logic held up at the county level. But at the block level, it starts to blur into something that feels less like risk pooling and more like… well, like punishing you for where you live. Luna: And that's where the ethics get thorny. Because where you live in the US is still heavily correlated with race and income. So an AI model trained on zip code level claim data can effectively reproduce redlining patterns. Lucas: That's exactly what a 2024 study from the Consumer Federation of America found. They analyzed rate filings in six states and discovered that, after controlling for driving record and credit history, drivers in majority-minority neighborhoods paid on average seventeen percent more than drivers in predominantly white neighborhoods with identical risk profiles. And the gap was growing as insurers adopted more granular AI models. Luna: So the AI is essentially learning that certain zip codes have higher claim rates, and because those zip codes overlap with historically redlined neighborhoods, the model perpetuates a cycle of higher premiums for people who already have less disposable income. Lucas: Right. And one of the tricky things is that insurers will argue they're not using race or income directly—they're using 'location' as a proxy. But a proxy that's this tight with demographics functionally does the same thing. The model doesn't know it's discriminating; it's just optimizing for prediction accuracy. But the outcome is discriminatory. Luna: This is the kind of episode where I'm glad we don't run ads. Quick honest thing—a handful of listeners chip in monthly through Buy Me a Coffee dot com slash fexingo, and that's literally what funds making episodes like this possible. No underwriters to worry about, no pressure to soften the take. If you've gotten something useful out of the show, it's a way to keep it ad-free and independent. Lucas: Yeah, it makes a real difference. And speaking of that independence—let's get back to what regulators are doing about these geographic pricing models. California has actually taken the strongest stance. The state's insurance commissioner issued a bulletin in late 2024 saying that using granular geographic data—anything more specific than a zip code—must be actuarially justified with a transparent model. In practice, that's slowed down some of the more aggressive AI deployments. Luna: Massachusetts has similar rules, right? They require insurers to file their rating algorithms for review, and they've rejected several that couldn't show the geographic factors weren't proxies for race. Lucas: Correct. And there's a bill in Congress—the Fair Auto Insurance Act—that would standardize these protections nationally. It would require insurers to demonstrate that any geographic rating factor is directly correlated with individual risk, not just area risk. And it would create a federal review board for insurance AI models. Luna: Has it got any chance of passing? Because I remember similar bills going nowhere in the past. Lucas: It's still early. The bill was introduced in February 2026, and it's stuck in committee. But the political climate is shifting. We've now had three major investigative reports—one from ProPublica, one from the Consumer Federation, and one from the New York Times—all documenting cases like Sharon's. That kind of pressure makes it harder for the industry to simply say 'trust us, the model is fair.' Luna: But the industry's counterargument is pretty straightforward: if you remove geographic data, you remove predictive power, and then good drivers in high-risk areas end up subsidizing bad drivers in low-risk areas. They'd say that's unfair in the other direction. Lucas: It's a legitimate tension. No model is perfect, and any choice involves trade-offs. But the question is whether the current approach is striking the right balance. When you have a driver like Sharon—perfect record, ten years with the same insurer, no claims—and her rate goes up twenty-two percent because of a crash that happened two blocks away, you have to ask: is the model actually measuring her risk, or is it just measuring the statistical noise of her neighborhood? And if it's the latter, then it's not really insurance anymore—it's just location-based price discrimination. Luna: And that gets to the heart of what insurance is supposed to be. The original idea is mutualization of risk, not individualization to the point where you're essentially rating people on things they can't control. Lucas: Exactly. And AI makes it possible to individualize to an extreme degree. But just because you can, doesn't mean you should. There's a difference between using AI to detect fraud or to reward safe driving behavior—which most people would agree is fair—and using it to carve up risk pools so finely that the people who need insurance most end up priced out. Luna: So what can someone like Sharon actually do? If she's stuck with a rate hike based on her zip code, what are her options? Lucas: First, shop around. Different insurers use different models, and one company's AI might weigh geographic factors more heavily than another's. There are brokers who specialize in helping people find policies that use more individual-based scoring, like telematics—where you plug a device into your car that tracks your actual driving. Second, if you're in a state with regulatory oversight, you can file a complaint with the insurance commissioner. In some states, that can trigger a review of the insurer's rating algorithm. And third, there's always the option of contacting your state legislator. The Fair Auto Insurance Act might be stalled in Congress, but several states are considering their own versions. Luna: It feels like this is one of those AI ethics issues that's flying under the radar compared to things like facial recognition or hiring algorithms. But it affects tens of millions of people, directly in their wallets. Lucas: Absolutely. And it's a perfect example of how AI can embed bias not through malicious intent, but through seemingly neutral optimization. The model doesn't know it's treating people differently because of where they live—it just knows that zip code 64109 has a higher claim rate than zip code 66212. But the reason for that difference might have more to do with historical inequality than with driver behavior. And until insurers are required to explain that difference, the AI will keep amplifying it. Luna: Do you think we'll see a tipping point? A case that forces a major change? Lucas: I think we might already be seeing it. Last month, a class-action lawsuit was filed in Illinois against one of the largest auto insurers, alleging that its geographic pricing model violates the state's insurance anti-discrimination law. If that case succeeds, it could set a precedent that forces every insurer to rethink how they use location data. And that, combined with regulatory pressure, could finally push the industry toward more transparent and fairer AI models. Luna: Well, I'll be watching that case. And Sharon, if you're listening—we see you. Thanks for sharing your story. Lucas: Yeah, and if anyone else has a similar experience, send it our way. These real-world examples are what make the abstract debate concrete. For now, I think the takeaway is this: AI in insurance isn't inherently bad, but when it becomes a black box that penalizes people for factors outside their control, we need to ask whether the pursuit of perfect prediction is worth the cost to fairness. Luna: And maybe the answer is that we don't need perfect prediction—we need good enough prediction that doesn't undermine the social function of insurance. Lucas: Exactly. That's the conversation we should be having.