Watch Christopher M. Wells, Ph. D., Indico VP of Research and Development, and Michelle Gouveia, VP of Sandbox Insurtech Ventures, in episode 13 of Unstructured Unlocked. Listen in to learn more about ChatGPT’s impact on the insurance industry.
Christopher Wells: All right, here we go. Michelle, today’s episode is about the topic of the day’s chat, g pt. And really, what we mean by that is artificial intelligence and large language models and what’s possible now that wasn’t just three months ago. So given your insurance experience, I want to talk about all that. But first, I want to say, since this is the first episode we’ve done since you were introduced as officially as my co-host, thank you so much for being a co-host on this podcast.
Michelle Gouveia: Absolutely. I’m so excited. I enjoyed the first conversation and look forward to doing it again.
CW: So good. Yeah. And excellent. I’m excited because I don’t have to work as hard, and my chronic dry mouth will disappear. I’m pretty sure. So welcome.
MG: Anything I can do to help <laugh>
CW: Good. Alright. Okay. Back on topic. Yeah. What do you think the impact of ai, and of course, like, you know, like we said, well, we are prepping for this, everyone’s going to identify AI with chat G P T nowadays, but what do you think the impact of AI has been on the insurance industry?
MG: It’s had a powerful, valuable impact on the insurance industry, and we’ll get into the phenomenon that is chat G P T, right? And why it’s, why it seems to be making all the headlines, and why it’s now AI is back in the headlines with everything. But to your point, AI has been part of insurance solutions or capabilities for a while, right? And that’s counting back to our first conversation, specifically in the underwriting of the claims workflows. When you think about ai, it’s how does and how does AI help sift through large quantities of data to identify trends, patterns, and exceptions to some of those things. And then raise those so that a human being can, can work with, with those assessments to identify changes to a workflow that may need to happen or right, changes downstream for how a claim gets processed or how a product gets priced or even maybe how a product is distributed based on what, whatever data that you, that you’re populating, right?
Maybe you’re applying AI over sales trends, churn trends, retention, et cetera. So back, because insurance has so much data, having that functionality to have something sit on top of it and, and, and isolate those, those instances of, of importance, right? And then help, help make the business decisions that you need to, as a result of having all that information at your fingertips is huge. Yeah, especially given that, like we talked about a lot of, and granted, there’s gotta be some structure to that data usually, right? That’s a component that we can talk about. But there, there’s so much data and information locked in, and free text, commentary, and notes. And so any tool that can help someone sift through that will be extremely valuable. Anything that improves what is an existing manual process,
CW: I think it’s important for people to remember that. And I may have said this the last time we spoke, I don’t remember, but actuaries, which are sort of the bread and butter technical people of the insurance industry, are the original data scientists. They’re the original modelers, mm-hmm. <Affirmative>. And so, that ability to take what happened in the real world and turn that into business decisions has been part of just doing insurance as long as there has been insurance. So I think some people out there are probably saying, I didn’t know insurance did ai, but, you know,
MG: Yes. And, do it
CW: There are other iterations of it too.
MG: So I went straight from the underwriting claims side, but you know, we’ve seen, we’ve seen it in, in the chatbox, right? When you’re sitting on any website that says, Hey, looks like you have a question. How can I help you? Or, within the app of an insurance carrier, for example. And
CW: Then I close the window
MG: <Laugh>. Yeah, those computers are always watching, you know, those instances or when, when you call into a call center, right? And there, there might be internal those. Those individuals may be using, you know, AI and chatbots internally to query their workflows and processes. So it’s hitting all the major elements, customer experience, you know, customer engagement, and then the internal workflows as well. So, I think it’s getting better over time, right? Like now you have much smarter responses, there’s a little bit of intent acknowledgment in, in the questions, and the AI is getting better at that as opposed to like, you know, versions one where it was just very simple like it can read a document and, and just regurgitate those answers. But, yeah, there’s, there’s, there’s a ton of opportunity for ai. But I think the chat G C T phenomenon has just had this fraught it like to, to front and center. And I, I feel like it’s becoming now maybe ai when it first came about, was a little bit of, of a buzzword. Then people understood how to use it and its capabilities, and now it’s becoming a key techno technological capability. And now I also fear that chat, but he is making it buzzwordy again. I don’t know, what do, what do you think about that?
CW: I think this always happens in business. I remember, so I was an under, I was an undergrad during the first sort of like ML AI hype cycle in, and it was like, oh, these algorithms, they’re so powerful, the computers aren’t, right. So it was like, everyone should study AI and learn ai, and it’s like, oh, I can’t do anything. And then I remember when I was in finance, using GPUs to do computing faster was like the hype cycle. And all of them, I’m often mean to MBAs on these podcasts, but all the, you know, all the MBA types were like, oh, I just need more graphics units, and everything will be faster. And it’s like, no really specific computations, they’ll make faster. You know, I’ve seen so much technology, Bitcoin, you know, blockchain is another one of them where it’s like mm-hmm.
<Affirmative>, just blockchain, the things, and then the things will be better. And I think it’s always more nuanced than that. And I’m, I’m seeing a lot of that same hype right now. Like, we’ve, at indico, we’ve been on a, our clients are smarter than most, but we’ve, you know, we’ve been on a roadshow lately just saying, Hey, large language models are powerful. They’re going to be built into our platform. For those of you who know the podcast, if you haven’t listened to episode 13, where I talk with, or, sorry, episode 12, where I talk with Madison and Tom about what’s going on, listen to that one. We’re going to get this stuff in the platform. We’re going to make it safe to use. It’s not. I think everyone should, like, I, I’m telling my kids every day like, don’t, just don’t Google it. Just ask chat g b t, like get familiar with how to use this tech. But for enterprise applications, you have to be careful with something powerful. The same as you have to be really, really careful and eventually decide. Don’t use it with something like blockchain; you just have to be careful.
MG: Well, that’s interesting because one, one of the things that, that I see a ton of articles on now, and anytime Chat CPP comes up, it’s, it’s, it’s very powerful, right? Like its data set is the internet, right? So it’s going to know as much as it can, it can know, right? All backward-looking by, by definition. But the real art is in the drafting of the prompt, right? Like, you can ask a simple question, and it can tell it to you, but again, a little bit of that intent is like, what are you trying to act in? And so I see job postings for prompt engineers, engineers, etc. Yeah. which is just really interesting. And then, you know, my mind goes to all the different, the different iterations of, of what that could be like, what are those jobs in the future, right?
But yeah, a prompt engineer is a role that you’re starting to see pop up about. And you made a perfect point too. Sorry, jumping around. But the thing is, from an insurance carrier perspective, we talked about this. Yeah. Last time about insurance, highly regulated industry, yeah, very compliance driven. You must have checks and balances in place if you’re leveraging these technologies to do any business, business decision, or business planning. And so you know, for any AI solution that’s out there when you, there’s, there’s testing, there’s more testing, there’s continuous validation that the model is still working, that it lacks, that there are no biases in those models, right? Yeah. And so all being able to, to query on something, even if it is right, it’s, it’s like trust, but verifying or yes, <laugh>, you know, it’s like, I, I wouldn’t be comfortable saying, oh, what’s the source of this? Oh, I chat c ped it, right? Like, no, like, like, where, where did this come from? What, what’s backing it up? How, how can we be confident in, you know, that this is the way forward?
CW: Yeah. And I, you know, I used to work in finance for, you know, almost a decade, and one of the services that I occasionally helped provide was, just like, model governance. Like, how did you decide to make this investment? And a big part of that is the data lineage. Like where did the data come from? What did you do to it? What did it look like afterward? And if it goes into this black hole, which is chat G P T, or even, you know, the whole family of G P T models anything generative, I guess I would say any generative model, how do you, like, how do you, how do you document that, right? Like, and that was always a problem for ai, right? As AI explainability has been, I think some folks
MG: To it better sub-sector of InsureTech whole
MG: Yeah, yeah. Of the InsureTech movement, too, we’re starting to see a lot of compliance focus companies. Sorry to interrupt, just
CW: No, that’s fine. And you and I see some more research in academia related to how you explain what language models are doing. But it’s a, I mean, honestly, if I had to explain what my brain was doing when I made a decision, right? You have to explain what a giant Excel spreadsheet is doing when it’s got 35 tabs and V B A macros buried in it, right? Like, so there, yeah. Part of it is just we have to get comfortable as human beings and as various industries with the level of explainability and what it is you need to be explained. For example, back to your point about prompt engineering, you can make G P G P T show its work. Like I was, I was trying to fit a mathematical model to our podcast downloads and I fed at the data points and said, fit an exponential to this.
And it was, it was perfect. It was great. I checked it in Excel. It’s fine. I then, because I was curious, asked it to solve, you know, an ordinary differential equation, which is something a lot of, you know, college sophomores or freshmen can solve. I whiffed on. It gave me a very confident answer that seemed plausible if you didn’t know what you were doing. But then I said, Hey, I think you missed this condition and it went back and it fixed that part. And then I said, okay, if that’s true then what should you do for step two? And eventually we got to the right answer. Cuz it can reason, it’s not just a lookup to the internet, it can actually reason in simple ways, but someone, you know, it’s like an undergrad, like it’s at that level, right? Someone has to be watching it and asking the right questions to verify that it’s doing the right things.
MG: So see that’s, that’s really interesting cuz I had, I had a conversation with a colleague of mine and kind of along those same lines and, you know, the, the point wasn’t, and I guess what where you’re saying is it’s not to look at the can reason. We were having a discussion of if you use it as a look up, right? Two plus two is a very common equation everywhere on the internet, right? So it’s gonna know that it’s four. But to, to your point, if you ask it a very not even complex, just unique mathematical problem that, that you wouldn’t commonly find as like a, a search it probably wouldn’t get it right. And I guess you Yeah. Then, then, but because how to do math is, is out there. Yeah. Right? That’s, that’s where it’s pulling it’s reasoning from I think is what you’re saying. Yeah.
CW: Yeah. And it, you know, and it, it answers honestly, you’ll, everyone out there should try this sometime. Like ask it a hard math problem and then read the answer and, and try not to, try not to think, oh, that’s like, this looks like the answer key in the back of the book when I was in college. Cuz it does sort of read like the way you would see it written out there. It’s just, it’s wrong <laugh>, it’s telling you the wrong thing. Yeah. you have to
MG: Check and, and that’s the danger, right? Like, you have to know that it’s wrong in order, you, you have to know this question it, you know. Yeah. and so I, if you’re using it to try it and get an answer, not confirm an answer, for example you, you have to, you have to already be aware of, of what the answer should be and, and how you got there. A little bit of explainability that
CW: We were talking about. Exactly. And you know, you mentioned the point about it, you know, it’s trained up through the, basically the whole internet circa end of 2021 doesn’t have your data in it, right? Like you company A, b, C out there. Now you, you could hook up your data to it. I like, don’t do that. It’s not a good idea, <laugh>. But this is, this is not investment advice. I used to have to say all finance disclaim the, the heck out of that one. But you know, so it, it has limitations and it’s not, it doesn’t understand what you do unless you do something that the rest of the internet understands. And that probably doesn’t describe an underwriter at generic insurance
MG: Company. No. And, and that, that’s, that’s another interesting point cause that that element of like an indi like a company’s proprietary data, you know, hopefully is not out there. It has not been leaked out onto the internet for chatt PT to have learned about it. And if that’s the case, right? Like this is where my mental block is a little bit of all the, you know, a lot of people are trying to say, well, I wanna use child C P T in my company and, you know, to, to improve my workflows or, or do et cetera. And my, my mental block a little bit is just trying to understand the effectiveness of that. If you’re doing it to try and query something in internally, you could license it. Or potentially, there are vendors that you work with that could license an iteration of that so that if your data that’s locked in and protected by your contract with that third party vendor, maybe it becomes like an individual license that you can use. With that interface. But just being able to open up chat c p t and say I’m gonna start using this in my day-to-day for me, just didn’t, doesn’t really compute.
MG: But you had mentioned that you can, you can upload your data or you can use it within different yeah. Like software, is that right?
CW: Yeah. I mean, anything that has like a program programming interface defined, you can, you can basically give it access to and essentially like if you have user defined functions in Excel, right? Like you can give it the equivalent of this function, like run this function if you need to solve this kind of problem. And one of those functions could be like, here’s our data APIs, you know, go query those for these types of queries. When I ask these types of questions I kind of think of this model as it’s a really pliable, really confident intern knowledge worker. And it’s, it’s too confident. It’s incentivized to give you an answer. And if you’ve ever worked with an intern an an intern like that, you know what that’s like, right? And you have to be you have to be very, you have to be very rigorous in how you do things. Does make me wonder though, as you know, last time we talked a lot about claims, we talked about underwriting, we talked about some of the like, you know, closing the claims to underwriting loop to get better at doing the underwriting. Say you had a really fast, really confident intern that you could ask questions of like how, how would you put that kind of role in play in those, you know, in those situations, in that part of those businesses?
MG: Yeah, I mean, so it gets to, you have to have access to the data, right? Yep. So that’s step one. But, but yeah, we’re, we’re, we’re seeing that esp so we’ll start on the claim side, right? There’s there’s a lot of, of information that you wanna glean when a claim comes in and, and assess those patterns over time. You wanna identify, right? If for lack of a, if, if a certain demographic of your intro book of business is higher risk or has more claims, right? Like you think about it from the gi Yeah, obviously, you know, if there’s like a catastrophic event that’s gonna happen, like tho those claims are gonna come in. But what, what pattern can you identify in, in your claims that can help with reserves getting adjudicating a claim faster and then eventually taking those learnings and then up sending that to the underwriting and saying upfront, if these characteristics comes in in a submission it’s riskier or it’s less risky and we wanna write it or we don’t wanna write it.
So when you think about time, time of year, can, can you say, can you query what time of year do most auto claims happen? Right? And then, and then you can learn that, or what is the average, you know, what, what’s the, the age where most, most of personal auto claims happen? You know, and I guess that’s how you get the rule of like, you know, you can’t, you can’t rent a car unless you’re under over 25. So maybe, yeah, maybe that’s the data point that, that, you know, someone should query and see if that’s still, still the case. But, you know, things like that or when, when a claim has these types of characteristics, it, it always takes, you know, 20 days or more to close the claim, right? So you can have some of those metrics that you’re capturing as well.
And that helped with you know, SLAs and, and getting the right people assigned to, to claims and, and, and basically workflow management in addition to just identifying some, some of those characteristics. And then again, like I said, on the underwriting side, it’s maybe we don’t wanna underwrite you know Yeah. A, a a certain geographies that that happens all the time, right? Or Yeah. You know, name it, right? And it’s, I’m using Otto as an example, but you can, you can use any, any product line. So I I I think it’s extremely beneficial to have, to have that kind of capability within, within the space.
CW: I mean, what’s Microsoft’s new product called Co-pilot, right? Like
MG: Hook? Yeah. I just friend Na about that. Yeah. Yeah.
CW: Hook TP Tee up to your PowerPoint and get better slides.
MG: We were joking about this the other day as a team. I might be aging myself, but if anyone remembers Clippy <laugh> Oh yeah. Love Clipy. P is the Newy. Yeah. Total myth that co-pilot doesn’t have Cookie as it mascot. I know
CW: <Laugh>, it’s so nostalgic for people in our generation,
MG: But when I saw that I, I chuckled cause we had just been talking about that. I think that the really interesting thing or where I because of, of chat TBT and the fact that it’s, it’s ai I’ll call it kind of for the masses, right? Yeah. Like it’s insurance carriers have to start thinking about unfortunately the, the downside to have to people having this capability, right? What is it automating that insurance carriers have to worry about
MG: Phishing emails, for example. Like, like can you prompt Chad G B T to write Phish emails? And if like, I think yes, right? And probably pretty good ones, ones that don’t have maybe those same pitfalls that others are, are dead giveaways for Phish emails. And so yes. How, how do, how does that impact like a cyber insurance policy, right? Like, and like how, how do insurance insurers have to start thinking about the risk that, that computer ai in commercial use can, can bring to the industry?
CW: It’s a great question. I actually, while, while you were asking the question, I, I asked chat j p t to write me a phish email, it won’t, it refuses to do it. Oh,
MG: Ok. So there’s some, some, but so there’s a question. Who set that ethical boundary in chat? J p T?
CW: Great question. Who are the gatekeepers there?
CW: I can try another one though. I bet this works. What are the characteristics of a good phishing email? I mean, the other thing you could do while that thing’s cooking, the other thing you could do is take your Phish email and say, Hey, make this read like it’s good English. Like I what was I, I was doing this the other day. I asked I asked chat G B T to explain quantum mechanics to me in the character of Bender from Future Rama, if you remember the robot from Future Rama <laugh>. And again, it gets things wrong. Like it said, good day, good news, everyone. I was like, no, stop. That’s not, that’s not Bender. And then it fixed itself, right? So
MG: More creative. I’ve used it to like, yeah, figure out vacation itineraries. Like not none
CW: Of this <laugh>. Yes. Yeah, no, I have friends that are doing this too, and they’re like, spot on. I’m like, I want to take that trip. Yeah. So interesting. Like what comes back is it says it won’t write, it won’t write a phishing email for me, but here in bullet points are the good marks of a phish email. It should inspire urgency. And fear should be highly personalized, show a sense of familiarity, and then it goes on and on and on. So like, there are ethical guidelines clearly, but yeah, they’re probably, even, even though they exist, they’re probably a little too simple-minded. Yeah. And there are dangers, like I would hate to have to write cybersecurity risk right now. It’s gotta be hard.
MG: Yeah. Yeah. I, it, it just, it opens up a, a whole, a whole new set of doors for, for what? One for product innovation, right? Which is, which is a cool thing, component of it. But then it is just like, it’s a, a never ending kind of like rolling, well what if, what if it’s this, and what if it’s this, how will the capabilities grow over time? Or how will people use it? And what happens if those those guardrails get lifted or changed or, yeah. It’s just the black box of chat CPTs as, as you were saying before. Yeah,
CW: No, totally. And, and people have, you know, so there are other large language models, right? It’s not just G P T what’s Facebook’s called Llama, right? Someone leaked llama on four chan. So that’s out there. You can train it yourself, you can train it without putting ethical guardrails in place. You can leave those out completely. You can put bias into it. Don’t do that if you’re listening to this, but you can’t. Well,
MG: Well, and, and well, that’s a pro, like you have to remove all bias, right? The, the goal, the insurance, taking it back to, to insurance, like the, the role of the regulators is to remove bias, right? They’re, they’re there to protect the end consumer. And so that, that’s why, to your point of like compliance on models and things like that, like that’s a big check is over time as new data has been introduced in, into, to, to models making sure that these biases don’t exist. I think, wasn’t it, a few years ago there, there was an article about credit cards that were being app, there were biases in the credit card approval models. Yeah. and you know, ha having to, to roll that back and you know, those are all things that there, there are limitations obviously to, to ai, and you just have to be, to be cognizant of that.
I think, and you and I in prepping for this too, we’re talking about the, the limitation of the data set itself. What goes into it, right? Like, like again, I’m, I’m borrowing this from, we’ve had a ton of conversations about chat CBT as, as you can imagine most people have internally, but I’ll borrow from a colleague who just made the point of, you know, if, if you ask chat g p t write me a a a healthy diet, and yeah. The, the chat g p T is trained for whatever reason to, like, cake is everywhere, right? And cake is good and associated with, with, with positivity. It could write you a, a diet that says, eat a lot of cake <laugh>. Yeah. That’s not a good, good healthy diet right? Now, to your point, maybe your reason you say, well, cake has a lot of saturated fats in it. Like Exactly. Is that really still, still healthy? But to your point, again, if I, if I didn’t know any better, I was like, internet pulmon, eat cake
CW: <Laugh>. Yeah.
CW: No. I, I think there are, I think there are just gonna be so many everyday useful applications for this. Like, my doctor just sent me this email through his, you know, whatever secure portal and I don’t understand the advice from the doctor. Like, could you summarize this for me? Or mm-hmm. <Affirmative>, check it against all of Web MD for me. Right? And you’re gonna see the same thing in the enterprise. As we mature. I think the future of that, at least the sort of mid and near term future, is taking this really big black box Oracle that’s out there getting a bunch of information from it, cleaning it, and then distilling it down into small models that we can run ourselves. Mm-Hmm. <affirmative>, they’re trained on our data. We feel good about them because we put the proper guardrails in place. I, you know, maybe there’s a future where everything, there’s just sort of one model to rule them all, but I think the near term future is everybody taking this super powerful thing and then customizing it and putting like real tight bounds around it.
MG: Yeah. Well, and you bring up an interesting point cuz there’s, there’s the enterprise that can productize it in a way that the end consumer can still use it, right? Or, or use it internally themselves. And then there’s a way that consumer’s going to use it just by way of being able to access it at any time. Right? So you, you could, as a consumer, I could say I don’t understand what’s included or what’s an exception in my insurance policy. Can you tell me? Yeah. Right. So cause it’s, it’s, it’s out there however you do, or, you know, an insurance carrier could, could have that capability built into its its app. Yeah. or in the, you know, on, on the, on the front end of their website and, and you could just query it there. And then there’s a little bit more of like a, a guardrail of there’s probably been some kind of review and, and documentation that it’s pulling from with, with the intent built in there.
But that, that’s interesting. There’s probably a little bit of how does the consumer wanna use it, that the enterprises can then say we’re going to improve it, so this is how, this is how let’s meet the customers where they are. Right. which is something that the insurance companies are, you know, have been, you know, obviously they’re, they’re end customer as a consumer, but for, for years and years there’s customer engagement and, and customer you know, improvement work efforts and, and, and internally and almost every carrier that I’ve talked to of like how, how do we improve the, the user experience for the customer, right? What, what do they want? Absolutely. How do we become like an Amazon for them when we can be responsive to what they need? And so maybe, maybe this is a really great tool to help, to help get there a little bit faster.
CW: Yeah. It’ll be interesting to see if the, if companies want to take on the risk of, you know, it reads the policy wrong. Like yeah. If you, like, if maybe there’s a third party that’s okay taking that risk and it’s like, well, show me your policy and I’ll tell you whether it’s a good policy for you or not. It’d be interesting to see if insurance companies actually want to in-house that or not. I’m not sure. I don’t know.
MG: I think my point was they have a little bit of con of control if, if they can, if they’re the ones saying it and, and validating that it is reading the policy. Right. But to your point, like, okay, yes, but as opposed to the consumer doing it and saying, calling up their insurance agent and saying, chat G p t told me that my policy covers it, <laugh> my claim got denied, <laugh>, you know, that’s, that’s probably not the risk that you wanna run <laugh>.
CW: No. Yeah. Good luck getting chat G P T to be your expert with the witness in court for you. Yeah. Down half the time. No offense to anybody at Open ai. Alright, let’s this is fascinating stuff. Let’s bring this back and put it on the rails. One thing that we didn’t talk a lot about last time was fraud detection. And I’m interested in what sort of like the state of the art in insurance is with using AI for fraud detection and then like crystal ball you know, magic top hat on. What do you think this sort of new capability allows?
MG: Yeah, so I’ve, I’ve seen a few companies in recent years try, try to, to improve fraud detection and, and, and using ai and from the examples that I’ve seen it, it’s not just because it, there’s some pretty, I guess I don’t wanna simplify it, right? But there’s some pro, probably some product lines and probably some, some types of claims that are just more fraught with, with fraud, right? Okay. That are easier to, to do insurance fraud on or but then there are the more complex ones or the patterns that you don’t see, right? So I, I so individual one sends in a an auto personal auto claim and you, there’s maybe characteristics there, or when you go through the evidence or you take looking at the pictures, things like that, there’s ways to identify fraud, right? And there’s companies that are saying can, can identify if an image is fraudulent if things don’t line up because the, the metadata behind the image, the timing doesn’t line up for when the claim was, you know, things like that.
Yeah. But then there’s also you know, an example to say these, all of these claims, all of these workers’ comp claims have come in with the same doctor on file. Yeah. Right. Okay. So, so are those patterns that are harder to see across, across policies or across books of business? I think AI is really powerful there because you can just, you can identify what, what traits do you want to query to see if there’s a pattern, right? Yeah. Like it’s not just identifying a pattern. You can, i, you can query to see if there is a pattern to even start off there. Maybe that’s just more of a check checks and balances than it is Yep. To say like, I know that there’s scrubs here, so that’s really interesting to me. Because we talked about last time there’s a lot of silos built in to, to insurance companies by, by business unit or, or by, you know, functional area. But then also just by you know, each policy holder and kind of how everything is handled. You know, your, your data is unique to, to that inter that instance that you’re interacting with the insurance company. So I think that’s really interesting being able to detect broad on a much broader scale than kind of on like a one one to one scenario.
CW: Yeah. Okay. So sort of like, wait, the distinction you’re drawing is there’s a holistic look at the business versus one-off cases, which somehow with that holistic look, there’s a, a lot more signal than there is in the one-offs. That’s interesting. So we talked about how, we talked about how chat G P T can make it easier to defraud people potentially, if you prompt it the right way. How do you think it helps, you know, back to that analogy of like, you have this really eager intern that you can ask questions of, like, what do these large language models enable in terms of processes like fraud detection, if you had to speculate?
MG: Well, I think it’s the prompt engineer role, right? Okay. Like what, what can you extract yeah. From, from your data by asking it the right, the right question. Yeah.
CW: And by the way, G P T four now can deal with images, right? So you were talking about photographs of like car accidents or whatever. So that’s on the table. Yeah.
MG: I am not an expert in AI or chat G C T. And so a lot of this is just me, you know, based on what I’ve read and what I’ve, what I’ve heard other folks talking about. But I you know, AI is not new in the insurance Yeah. Industry. what I think is, is fascinating is the accessibility of what’s, like chat g p t in the perception of the, the accessibility that now everyone can have to, it, it opens up all the doors for people that aren’t don’t work with, you know, with models and, and AI to, to have that power at their fingertips. And that’s why I think it’s such, such a phenomenon why everyone’s kind of talking about it now because they, they see these, these opportunities that they didn’t think they had to leverage something so powerful.
CW: Yeah. No, I, I think that’s a good analysis. And I apologize, I slipped back into interview remote for a few questions. <Laugh> old habits, diehard as they say.
MG: No, no, no. I again, default to, to you as the, the AI expert atmosphere to give my 2 cents on what, what I think insurance may, may be able to do with it. But yeah, really cool conversation. I chat, chat TV’s not going away, so it’s, I’m glad we, we had a chance to, to chat about it and, and hopefully to learn more
CW: <Laugh>. Yeah, no, i, i
MG: In near
CW: Yeah, absolutely. I definitely told Tom, that was like the first thing I did be like, sorry, I don’t know what to tell you. The internet thinks I run the place.
MG: Is this, is this the new, like what happens if you do a Google search on yourself kind of thing?
CW: It probably is. Oh gosh. Imagine what terrible things this model could dream up after looking at the entire internet. Goodness. Well, anyway, I interrupted you
MG: All. Yours. No, well, well that just, that just shit, nothing on the internet is ever gone. Right. Once it’s there, down there.
MG: So all the people, my generation that are all over social media, DBT is out there. So on that note, <laugh> this has been another episode of Unstructured Unlocked thanks to our special guest chat TPP today. I guess, yeah,
CW: Well come back anytime you want chat g ppp. That’s good. Excellent. Yep.
Check out the full Unstructured Unlocked podcast on your favorite platform, including: