Revisiting Early Misconceptions About Artificial Intelligence

Speaker A: Mutual insurers are built differently. Rooted in their communities. They have earned trust with agents and policyholders because they are committed to being there for their customers no matter what. That's why we built the Mutual Group to strengthen what makes Mutual special and accelerate what's possible. As a member based insurance services platform, we help mutuals modernize, grow and compete at scale while they keep their brand, culture and independence. The Mutual Group dedicated to mutuals partnering for the future.

Speaker B: Welcome to Insurance Uncovered, the first podcast to bring you insurance news and an inside perspective from thought leaders in the property casualty insurance industry. Insurance Uncovered, um, is a product by the national association of Mutual Insurance Companies and is sponsored this week by Holborn. We're continuing our series of revisiting some of the most listened to unscripted segments of 2025. This week we're returning to a topic that everyone is talking about and will be talking about for the foreseeable future, Artificial intelligence. Let's listen to Namix CEO Neal Ulridge chat with NAMICS resident AI expert Lindsey Clarkowski Stefani about early misconceptions around AI.

Speaker C: No matter where you go, it seems everyone in really every context is talking about artificial intelligence these days. Uh, it's certainly a hot topic in our membership here at NAMIC as well, bubbling regulatory and legislative interest in the topic. Uh, what you may not know, well, some of you certainly do know, but some of maybe all the listeners don't, is that NAMIC has actually hired a, uh, vice president of Artificial Intelligence and Data Science and Cyber in the form of Lindsay Clarkowski. Lindsay is an attorney. She's been with us. I don't know, Lindsay, what, a little over a year now, right?

Speaker D: Maybe 18 months, about a year and a half now.

Speaker C: Yeah, Yeah, I thought so. And Lindsey covers this topic for us very well. Those of you that have had interactions with Lindsey know she the powerhouse on this topic. And so we are going to talk about the whole artificial intelligence landscape as it relates to insurance regulation and what members can expect and what we anticipate here kind of at every level. So Lindsay, thanks for joining me today on the podcast.

Speaker D: Thank you for having me. I love this topic. Very near and dear to my heart.

Speaker C: Yes, it is. And you are one of the few people that really have some expertise on it. So we're glad to have you here. So why don't we just start kind of at the beginning, uh, you know, talk about the, you know, little bit of history on this topic as it relates on the regulatory side. It's sort of a short history it's probably more in the future than it is in the past. But talk a little bit about where we are now and what we're facing.

Speaker D: You bet. So you're right, the policy history relative to AI is very recent and relatively young. Um, it only started to garner a lot of attention over the last two and a half years or so. But the overall themes that we see popping up in the AI policy space actually started showing up and popping up a little earlier than the widespread popularity and kind of this recent mainstream explosion that AI is experiencing. So when I talk about the advancement of AI policy, I like to think of that advancement as having three distinct eras that have tended to coincide with advancement in technology generally. So the first of those eras I like to term, um, it being the pre generative AI era. And I put this era in the 2000s to 2010s where the tech advancement capturing everyone's attention was the onset and availability of what we call big, big data. And big data is often characterized by a large amount of data, the wide range of data types and then the speed at which the data is generated and shared. And if we think about the 2000s to 2010s, the Internet and computer advancements throughout that time really helped big data take off. Now from an insurance policy perspective, the focus during this time was less on a piece of technology than it was on an insurer's use of new or non traditional data sets that were becoming available through the rise of this Big data. Insurance as we know is in many ways ah, a data driven industry. And the availability of big data held and still holds great prospect, um, particularly in the way of risk based pricing through increasing precision in underwriting and rating. And not only does expanded availability of data elements help increase that precision, but if you start applying predictive models or algorithms on top of big data data sets, you can start to identify patterns to help better inform risk rating and underwriting decisions. So there's lots of benefit to be gained there. And insurance really dove in during this time. But the regulators and the lawmakers approached all of this with some caution. And they were largely. They do, they do, and understandably so. But um, with this one in particular, their concern was um, this theory that some of the new or non traditional underwriting data may correlate more strongly with a particular protected class, such that the protected classes are disproportionately impacted relative to charged premiums or otherwise. So as a result we started to see the term and idea proxy discrimination be thrown around. Um, we saw this at the end coil level at the NAIC level. And we saw New York Department and Colorado Department take some action in this space with a circular letter and a proposed law, respectively. So this pre generative AI era was one from a policy perspective that was really focused less on a piece of technology itself and more on the expanded use of big data and its perceived impacts. Now, if we move into what I, um, call my second era, this is when generative AI hits mainstream. So we are in November 2022. Generative AI and the release of OpenAI's ChatGPT3 was really the first time that AI as a technology became widely available and widely accessible to the general public. Now, AI as a technology and as a concept has been around since the 1950s, but generative AI was a turning point in both advancement and availability. It particularly brought about what I call the democratization of the technology, if you will. And what I mean by that is anyone who has access to the Internet can use generative AI and create new content without needing to have any technical expertise. The main hallmark or differentiator of generative AI is that it's trained on large amounts of data by scraping the Internet, and it can generate new content from a prompt. So I don't know if you've played around on ChatGPT or not, Neil, or, uh, other large language models out there.

Speaker C: Yeah, not much really, but a little bit.

Speaker D: But, you know, it's fun. I mean, by way of an example, you could get on there and tell it to write me a poem about insurance in the style and prose of Mickey Mouse, and it will generate an output based on that prompt. So this, this widespread availability and the large jump in capability brought with it amplified attention from all levels of government and all business sectors. And what we saw from the end of 2022-2023, largely policy and lawmaking bodies trying to wrap their arms around what the technology was and what risks there may be that might require new legislative or regulatory efforts. So this second era of generative AI hitting mainstream is one that I characterize as an educational era. There was a lot of swirling, a lot of learning, but not much concrete action. But that started to change with this third era. And the third era is the one that I like to say we're in right now. And so far, based on what we've seen, um, I'd characterize this era as one that's focused on ideas of algorithmic bias and algorithmic discrimination. There is in the background, definitely, a battle of authority over AI regulation playing out. But the resounding theme Coming out of all of that so far has been on AI, explainability, bias, disclosures, discrimination, black box concerns, you name it, and that's across the board. That's not only specific to insurance, though. Insurance hasn't been immune from these types of policy conversations. But by and large, there is this focus of this concern and idea over algorithmic bias. That's a lot of what we saw in 2024, and it's a lot of what we're seeing in the first week and a half of this new year and what we can continue to expect.

Speaker C: Yeah, well, that's a great, you know, sort of primer as to where we are now. And so we have a working group here at NAMIC that you're heavily involved in.

Speaker D: We do.

Speaker C: It is, uh, aptly named the Artificial Intelligence Working Group. We, uh, tend to call things what they are around here. Uh, and so that group has been working at the federal level, at the state level, at the NAIC level. I, uh, don't know if you want to spend just a minute there on kind of its focus and. Or maybe each area where we see particular problems you might want to highlight.

Speaker D: Absolutely. So our AI Working group is, um, a working group that I lead here at namec. And it is comprised of many, many, many of our members who generously devote their time once a month with me to discuss kind of the hot topics of the day in the AI space, which, as you can imagine, is a very dynamic space. Space. I think I can best illustrate, um, the work that we do with what we were Busy doing in 2024. So 2024 kept us very busy. Um, I'll start with maybe the federal side. And over on the federal side, there was a lot of action from the administration last year, but no real concrete action from Congress, though there were many proposals and many task forces popping up all over the place. The administration issued everything from a White House blueprint on AI to a sweeping executive order. And we saw various executive agencies take their own action. Particularly important to insurance was the action that treasury was taking and what the Federal Insurance Office was doing with their interest in AI in, I want to say, maybe mid 2024, treasury issued a request for information on UM risks and opportunities of AI in services. And NAMICS. AI working group was very involved in that, in, um, drafting a response to submit because that RFI contained a number of questions specific to insurance and FIO. Subsequently, in the fall of 2024, held an AI and insurance roundtable, which Namec also participated in with the help of our AI working group member Rachel Drade Rice from Next Insurance.

Speaker C: Rachel was on the podcast talking about this.

Speaker D: Yeah, she was. And I was going to say we're very thankful for Rachel's participation there. And if anybody hasn't checked out that interview, you definitely should. Rachel does a great job of walking through AI in insurance, coming from an on the ground perspective and talking through um, the issues that were brought up in that roundtable. So we were very involved from everything that the treasury and FIO was doing on the federal, federal side. And then over on the Congressional side, Congress seemed to agree that it wanted to do something in the way of AI legislation. But there was this divide over what exactly that looked like. For instance, should it be an omnibus, um, type of bill that covers all businesses equally, or should it be sector specific and then fill in gaps where necessary? So from the AI working group perspective, we continued to raise those bills and ideas as we saw them and kept engaged, kept our ear to the ground on those various proposals that popped up. Luckily, nothing gained much traction. There were a lot of questionable ideas as it pertains to insurance. So we definitely, um, kept our eyes on those. Now on the state side, we saw a little bit more concrete action in 2024 that kept the working group busy. The states got impatient in the wake of inaction from Congress and we saw widespread introduction of bills focused on the concept of high risk AI and algorithmic discrimination. These bills were pretty far reaching. Um, they applied to developers and deployers of high risk AI. And high risk AI. Yeah, high risk AI is defined as AI that's used to make or assist in making a consequential decision. And consequential decision is defined to mean decisions affecting access to a number of things. But importantly for us, it's often defined to include access, availability or pricing of insurance. So that's where this became, um, a huge item of interest for us and of work for us in the working group. Because for these high risk uses of AI, which included use of AI and insurance access decisions, these bills prohibit any resulting algorithmic discrimination, which to generalize is often defined very similarly to disparate impact. So Colorado signed one of these bills into law last year and Connecticut got very close. Senator Maroney from Connecticut, um, then started up a multi state working group for legislators from. I believe it's almost all jurisdictions now to develop model AI legislation with this focus on high risk AI. So there were. Although Colorado was the only state that signed one of these bills into law, there were a growing number of states introducing things like this. So, um, The AI working group, um, drafted up comments submitted, worked with NAMICS RVP's to get on the ground and really talk to these legislators that were introducing ideas like this. Finally, I'll touch on the state insurance department side of things and what was coming out of the NAIC. This activity largely surrounded, um, adoption of the NAIC AI model bulletin. So in 2024, we saw 21 state insurance departments adopt the NAIC Model Bulletin on the use of AI and insurance. This bulletin focuses on expectations for insurers to develop AI systems governance and AI systems testing programs, and that those programs should include testing of outcomes for bias. So there's that algorithmic bias theme again. And that was a point that, um, as a working group and namec, we got on the ground and really tried to talk through with the state insurance departments, um, how that metric for judging insurer conduct is brand new and conflicts with the existing unfair discrimination standard that is already here in the insurance codes. So that kept us very busy and we'll continue to see more of that in 2025 as well. And then we had Colorado's department and New York's department taking totally separate approaches. Colorado continued its rulemaking process for its law on the use of external consumer data and information sources, which largely prohibits resulting disproportionate negative outcomes on protected classes and insurance practices. And then we had New York issue a circular letter number seven, which at a high level includes required input and output testing designed to assess data correlation with protected classes and to prohibit disparate impact on protected classes resulting from the use of AI or the use of external consumer data in underwriting and pricing. So as a working group, we had a number of things we were keeping tabs on and engaging very thoroughly on. And as we spoke about earlier with those eras of AI policy, what we saw is this very large focus on algorithmic bias across the board.

Speaker C: Yeah, no doubt. And that raises all sorts of challenges for us as an industry. Obviously, you know, nobody wants to have something that has algorithmic bias, but proving it or proving the negative, very challenging to do in a realistic fashion and certainly, uh, something that we're wrestling with. I anticipate this is going to be with us for a while. Would you agree?

Speaker D: I definitely agree. I think, you know, from the federal side of things, with the incoming administration, I think we're going to see a little bit of a step back from the focus on algorithmic bias. This, in this incoming administration.

Speaker C: On the federal side, anyway.

Speaker D: On the federal side? Yes, definitely on the federal side, um, the incoming administration There has this focus, more so on national security and keeping America competitive as it relates to AI. But that's not to say we're still not going to see, um, numerous proposals coming out of Congress. So I'm sure we'll be playing a lot of whack a mole there still. Um, and then from the state side of things, the states are going to keep us busy this year, especially, especially with that turn of administration at the federal level and then the uncertainty of what Congress may or may not agree on. States are gonna go all in and taking their own action and we're gonna see a continuation and really proliferation, I'd say, of these high risk AI bills, um, that include this focus on algorithmic bias. We've already started seeing it in the first few weeks of the new year. Um, over in California, we expect Bower Khan to reintroduce Assembly Bill 2930, which was on this. In Connecticut, we're, um, expecting Senator Maroney to reintroduce Senate Bill 2. We're expecting something out of Maryland here soon as well as New York. Um, in Texas, we got a pre file of the Texas Responsible AI Governance Act. We have a carryover in Virginia with Delegate Maldonado's bill. And then we are expecting reintroduction of House Bill 710 in Vermont from last year. And all of these are a flavor of that high risk AI type of bill. And that's even just what we know in the first week and a half of this year.

Speaker C: Yeah. And I imagine that list is going to grow here as we continue to wrestle with this. Well, listen, Lindsey, I know we have a white paper coming out on this. Don't we do another one or a different version of one?

Speaker D: We do. So this is a brand new white paper that I've been working on and I've been trying to focus it on the themes that we've been seeing out of 2024 and what we're going to see in 2025. So this white paper talks all about these conversations and concerns over bias over use of big data and AI and how some of these concepts, or really a lot of these concepts, are misapplied when we're talking about them in the context of insurance, given insurance's very unique function, its unique pricing and its unique foundation. So this white paper that I have coming out soon goes through a lot of these misconceptions and aims to educate on all of these points, which should be very helpful and very timely as we work our way through 2025. Um, it goes through and explains what big data and AI is and how both of those are used or can be used in insurance. Risk based pricing. It goes through and describes the function of what risk based pricing is in insurance and how it works. And then it dives in and walks through the five of the more common misconceptions that have been swirling about in the past year or two. So first and foremost, why insurance products are different from any other consumer product. Second, looking at those concerns that are being raised over creation of a risk pool of one with increasing precision. Um, third, looking at what does fairness in insurance pricing mean and what should it mean. Fourth, looking at how the concepts of bias and disparate impact are incongruent with the risk based nature of insurance. And rounding it out, um, the paper looks at the limitations on outcomes testing, as you alluded to earlier, relative to insurance pricing models.

Speaker C: That's great. Well, I recommend everybody I know the paper will be excellent. It'll serve as a great resource not only for our advocacy, but also for the membership to use in their own, uh, working with policymakers. We have a, we're sort of at the beginning of the beginning of this, you know, saga as it relates to the use of artificial intelligence and insurance and its intersection with risk based pricing and all of the myriad of issues that are in front of us here. Um, it kind of makes me long for an old fashioned credit scoring debate, um, that we used to have 20 years ago that I still had the scars from. Um, but you know, it's an entirely different set of complexities here from a policy making perspective. And so, Lindsey, thanks for the explanation today. Thanks for all your work.

Speaker D: Absolutely.

Speaker C: Membership, on this topic, you've really come in at the right time and have, uh, really already made left a mark here in terms of your work at namic. So we really appreciate that you're an excellent resource for the membership, but, uh, a lot more to come here. This won't be the last time I think we talk about this on the podcast, but thanks so far and thanks for the update today.

Speaker D: Definitely. Thanks so much, Neil.

Speaker C: Uh, sure thing.

Speaker B: That's all for this week's Insurance Uncovered. Thank you to Holborn for sponsoring this episode. Until next time, have a wonderful day.