Fireside Chat: AI Hype, Fraud Reality, and the Fixes That Actually Work

Most conversations about fraud either overcomplicate the tech or overlook what’s happening on the ground. This one skips the noise and gets real.

Join Darren Thomas, Director of Data Solutions at Point Predictive, Justin Davis, Vice President of Product at Point Predictive, and Nathan George, Head of North America Partnerships at GDS Link, for a fireside chat built around what lenders are really dealing with today—from synthetic identities and fake documents to how fraudsters are now using AI better than some lenders.

You’ll get a clear breakdown of how machine learning fits into daily fraud prevention, why generative AI creates new risks, and what to watch for in digital and face-to-face channels. The conversation also covers how to test fraud solutions effectively, reduce friction, and measure what’s working without slowing down your process or frustrating your team.

Listen to the episode on Apple Podcasts, Spotify, or YouTube.

Episode Transcript

Justin Davis: For sure.

Nathan George: Thank you for joining us today. Today we're going to be skipping the slideware and as much as possible, all the normal buzzwords that you hear flying across the ether these days. And instead our goal is to focus on real world fraud issues that you guys are all facing right now. We want to make this as helpful and practical as possible for you. I'm Nathan George, the head of North America Partnerships at GDS Link and I'm going to be the host for today's Fireside Chat. I have with me today two leaders in the fraud prevention space. Justin and Darren have been fighting the good fight since the early days of basic static rules. And today we're here to talk more about AI-fueled threats that lenders are having to contend with on a regular basis. First up we've got Justin Davis. He's the Vice President of Product at Point Predictive. Justin has a background in forensic accounting and he's been in the trenches helping clients for over a decade, saving them over a million dollars and $100 million. Sorry. In fraud losses. Justin's led fraud strategy initiatives. He's built cross functional teams and is a great go to expert in the field outside of the daily combat with the fraudsters. He also serves on the board of directors for his local church and supporting mission work and community initiatives. Also joining us from Point Predictive is Darren Thomas. He is the Director of Data Solutions and is the guy behind the models leading the data science and engineering teams while they help lenders fast track the good applicants and flag the risky ones and do it at speed. Darren brings a deep technical perspective and knows how to make machine learning work in the real world. And we'll be leaning on him heavily for the tech parts of the conversation today. As I mentioned earlier, what we're hoping to cover are both the challenges that we see brought into this fraud free for all by AI, but also the opportunities to use AI to combat them and further our business goals. We're going to be talking about how AI is being used to fight fraud, how it's being used to commit fraud, cover some practical advice to help you learn, test and execute new strategies that will help you combat some of this. So let's get into it. Our first section, we're going to talk about defining AI and what it is and what it isn't. Just be careful. We don't want to cause hallucinations. We're going to start with a quick poll question and we're testing new technology at the same time. So this question is which of the following would you consider an actual use of AI in lending. We'll give you guys a few seconds to answer. All right. So far it seems most of our results are falling into the machine learning models bucket. It looks like we might have a couple in the rules based system, the expert systems bucket. I'm going to go ahead and close the poll out. Darren, give us the answer.

Darren Thomas: Sure. I think there's two correct answers here. You could say two and a half if you wanted to go back a couple decades I'd say for sure. The ML models.

Nathan George: Right.

Darren Thomas: That was clearly the most common answer chosen. That's what we're going to be talking a lot about today. That's really what we mean when we mean artificial intelligence is machine learning. It's a very specific type of artificial intelligence and it's a powerful one. The other one is the hybrid model stacking. All that means is throwing together two different machine learning models together to make a prediction. If you stack models along with rules, that is still a form of machine learning. I'll touch on the half one which I'll say is the scorecard approach is actually the early days of AI automating decisions using computers. A couple there to choose from. I really think machine learning and the hybrid are the two correct answers.

Nathan George: Excellent. Thanks, Darren. So we hear AI thrown around a lot. There's obviously many different kinds, many different ways of doing it and hybrid models where you're combining old methods and new together. When it comes to fraud in particular and lending, what are some of the most common misunderstandings you see with respect to AI and the usage of AI? Yeah, for sure.

Darren Thomas: So what I'll start with doing is just saying what AI really is in lending, like you said. So at the end of the day it's using data science to help identify patterns to make smarter decisions or automate decisions entirely. Right. Some parts of rule based logic and expert system scorecards as we were talking about has been the norm historically. Right. So a simple rule would be, you know, if someone has below a certain credit score, decline them. If they have above a certain credit score, approve them. Right. That's a very simplistic view of looking at an individual. And there could be other rules you throw in there. Right. That's just keeping it very simple but it's not actually painting the full picture. I think what AI and machine learning really do is actually helps paint a better picture of that individual to identify risk. So you know, there's, there's multiple approaches that we can use. We can use, you know, an ML model to help predict the chance of someone defaulting on a loan, that, that's what AutoPass would be quite predictive or really just detecting fraud in the individual. So that would be a BorrowerCheck. Right. So rather than relying on these rule based systems, you can look at large amounts of data to make an actual prediction. Right.

Nathan George: Is there a particular thing you can call out that would like, where do AI models really give a big advantage over that older system? Obviously you've got a lot of data, a lot of complex data, but when you start to look at it, you know, you see things like velocity of applications from a certain individual or you might look at network intelligence from, hey, this is a, you know, an IP address from somewhere thousands of miles away in another country. Are there other areas where I can really find signals that might indicate fraud that, you know, wouldn't, may not make common sense just to the average person looking at it?

Darren Thomas: Yeah, I think there's a lot of those examples and we always touch on the human element combined with the data science element. Right. It's not just a data science model. There's a human element that should be interpretable and explainable. On that related note, I'd say when it comes to something like income and income misrepresentation and occupation misrepresentation, fraudsters are getting better and they know the rule based systems that exist. Right. So where data science and machine learning have a good path forward is they can adapt over time. Right. You can feed it new data to recognize new trends that the, that the human, the human rules might not be catching or the computer system rules that you have might be catching. So I think it's the ability to evolve in real time that is extremely helpful with machine learning.

Justin Davis: I think another way to look at it too is in lending. You have underwriters, right? And as you bring on underwriters, they're going to get better over time the more applications they look at. But there's a limit to how many applications that they're going to be able to see. Right. And they're constantly feeding themselves when they're viewing it. What's good, what's bad, what's good, what's bad. But when it comes to machine learning, you're able to essentially train an underwriter with thousands, millions, hundreds of millions of applications so that it's much more accurate, it knows what's good and what's bad across a much larger data set rather than just the individual applications that that one underwriter might be looking at.

Nathan George: That's a great way of putting it. So in addition to dealing with complex data sets and you know, creating this static model of behavior that might indicate fraud. You also get the benefit of that real time trending things that change. As fraudsters try to adapt their attack methods to get past the static filters you've put in place, AI is helping catch up. That kind of brings us to the next big topic is how fraudsters are using AI to commit fraud. I just know personally with some of our clients that we've spoken with, there is a lot of change and most of it seems to be coming from the extra ability AI, the extra capabilities that AI can give to fraudsters to do bad things. Can you guys share like what are some of the top two or three new types of fraud or fraud that's grown more than others that is being fueled by AI capabilities being put in the hands of some of these fraud rings?

Justin Davis: I'll kind of hit on this in a couple of different ways. So when it comes to fraudsters being able to leverage AI more at scale, there is the fraud vectors in lending and then there's also just more broadly across banking, right? With financial institutions and scams, AI has allowed fraudsters to go absolutely buck wild with scams, right? With scamming people out of their hard earned money in lending, you're going to fall into more so of like the, the document misrepresentation, right. We think of like from, for, for what we do on the income side. Majority of lenders we talk to, they don't, they can't trust any of the pay stubs that they receive anymore because it's, it's easier to create a pay stub online than it's faster to do that than it is to actually find my pay stub, download it and then submit it. Right.

Nathan George: Is it as easy as, hey ChatGPT, make a pay stub that shows I make this much a year and then you got one that looks legit and

Justin Davis: put it in, put it in an ADP format and there are AI generator, like pay stub generator templates that you can go on and you can just request it and it's, it'll spit it out in a matter of seconds from like identity verification, liveness checks. There's software now that fraudsters will use where they can, they can put a face on top of them that, that'll change with it, right? And so instead of just kind of like a, where it used to be maybe a, a fabricated license where they just print a, a picture on top and you have to determine if it's a, if it's, you know, a counterfeit or not. Now, even when you do the, the selfie checks and the liveness checks, I can place a face on top of me within that phone, within whatever I'm using, and that face will change with my head so it looks like whatever is on the driver's license. But even in the scam world,

Darren Thomas: I'll

Justin Davis: actually post a link in the chat where people can. Our co-founder, Frank McKenna has his own blog called Frank on Fraud. And he posts stuff weekly about what's going on in fraud and how AI is being leveraged. But this was an interesting one that it's called Super AI and it's a Telegram service. So the, you know, the Telegram app that anybody can download. Telegram is, is used heavily by fraudsters for communication and for even just selling information. Super AI is a channel that was created to kind of mimic like a ChatGPT, but there's 81,000 members of this, this one channel on Telegram. And it's, you just put in a prompt just like you would in a chatbot or a Claude or a Gemini, whatever you use, and it can spit out whatever you're looking for. And so frauds. But it's been trained specifically to do certain things, like for romance scams, you know, put a picture and help me compliment this person. Or maybe I want to do a face swap. I'm going to give you my face. I'm going to give you my target image. And now just put me into that target image. It'll create documents for you, it'll do automatic translations. So fraudsters can scam people in multiple different countries in their native tongue. And it'll also do grammar checks. So what you. We've all gotten text messages from people saying, you know, hey, UPS. Actually, I got one this morning saying my UPS package couldn't be delivered because there was no signature. Please go here to, you know, request a new date. Well, there was July was spelled J I L Y. And I knew it was not right. Well, they can go to this and it will do all the grammar and spell checks for you, and all you have to do is push it out. And so it's just allowed fraudsters to conduct fraud at scale where before they. They had to either, you know, do it themselves one by one by one, or you look at places like, you know, in the Philippines or Cambodia where they'll have like camps and housing units just dedicated to have people that just do it all day long. AI allows it one person to do the work that, you know, they would have had to have thousands of people do.

Nathan George: So that's like an actual job position you can apply for it sounds like in some places.

Justin Davis: Well it's not. They don't apply for it. A lot of them are actually trafficked, forced into it.

Nathan George: Right. Yeah. Well I, I'm glad you put the Frank on Fraud link up there. For those of you who are listening that haven't signed up for the Frank on Fraud newsletter, it is well worth it. It's a great plug. It's one of my favorite newsletters that I actually read and don't just put into the circular file every week really always at least one or two interesting stories that are great anecdotes in conversation that week. So highly recommended. One of the things I notice you guys we're going to put a plug in at the end for the fraud report that Point Predictive put out as I was reading through it, one thing that really stood out to me. Well one, there's a lot of fraud schemes that I hadn't heard about before, hadn't thought much about in terms of how AI is helping them be even more effective. But one of the things that I noticed was the credit washing and synthetic fraud trend that in the auto lending space in the US made up 45% of all the fraud. I think if I wrote that down correctly. Could you maybe walk us through an example of credit washing and synthetic fraud and what that looks like and then I want to follow it up with first party fraud because that's another interesting one too.

Justin Davis: Sure. So when looking through that a lot of what we're seeing is more synthetics on the, on the first party side which when people first heard about synthetic fraud, you know, say back in 2015, 2016, it really was more organized rings that were just starting to create identities within the bureaus. Well now it's gotten much, much easier to do that to utilize what's called a CPN, a credit privacy number. It's sold as like it's a legitimate SSN that can be used for or an alternative, a government like legal alternative to an SSN. But it's not, it's a stolen SSN. Most of them are just children's socials that are stolen and, and reused and these can be bought. I mean you go to legalcpn.com and they have claimed to have sold a million different CPNs since they were incorporated and their cheapest package is 80 bucks. So million CPNs at $80 they made a good bit amount of money. But even with say social media, you can go on TikTok, Instagram, even YouTube Shorts. And people are teaching others how easily commit both synthetics using a CPN to create a new identity. But then also even wash credit. And so credit washing is really just the systematic disputing of trade lines as identity theft. Right. So I say I'm a 580 credit score, I have 10 trade lines and five of those are derogatory, either delinquent or charged off. Well, what I'm going to do is I'm going to go to ChatGPT, I'm going to have it create me a letter that I can then send to the bureaus and say I'm a victim of identity theft. I'm disputing all of these trades through the e-OSCAR program. All the bureaus are going to send that out to every financial institution. And because the FIs are regulated under and they have, they have to go to the FCRA, the FCRA says within 30 days of dispute they have to investigate and they have to respond back. And so if they don't, then they're out of compliance and you win by default.

Nathan George: Right. It gets removed from your report.

Justin Davis: Exactly. And in 2017 the FTC had removed a key piece of claiming identity theft. So prior to 2017 you had to have two things. You had to have an affidavit of identity theft and you had to have a police report that was accompanying that or it could be turned away. 2017 they removed the police report. And so ever since then, if you go to the FTC, they remove or they send out a report every year called the Consumer Sentinel Network Data Book. And it's a great read for anybody that wants to understand fraud trends through the FTC is identity theft claims skyrocketed. 2017 to today it used to be about a 10 to 11% year over year growth over a decade prior to 2017 and now it's more upwards around the 60 to 70% growth year over year. And it's just because it's so easy to commit. There's really no teeth to it because you don't have to go into the police department to file a report. And anybody can do it. And think 2017, maybe a bank or credit union, some FI is getting a couple of identity theft claims a month, a week, depending on their size. Well now they're getting tens, fifty, a hundred, but they have the same staff. How are you able to investigate that level of identity theft? When it used to take you 30 days to do five or 10, now you're doing 30 days over 100. It's like the rubber stamp starts to come out. And once the rubber stamp comes out, all of those derogatory trades are removed from the credit bureau. And now that 580 is now a 750, they still behave like the 580, but you're going to approve them and you're going to give them everything they want because they're that 750.

Nathan George: Right.

Justin Davis: And so that's being done now at scale. Social media makes it prevalent and easy and teaches people to do it and then say like a ChatGPT makes it very easy to just create a form and send it off.

Nathan George: Yeah, that reminds me of the, I mean basically it was check fraud when they had the free cash thing that went out on the Telegram channels about going to the ATM and they were basically exploiting a gap and when your account got credited cash and they hadn't counted the check as fraud yet, I don't know if you remember that. Was that Chase? I think that's Chase, yeah. Yeah. But that was propagated by, you know, TikTok and Telegram and all these online channels. And you know, the slowest mover is the one that takes the hit. Right. The one that closes the gap at the slowest speed is the one that's going to get it. You know, we had a, we had a similar client experience where apparently a ring of fraudsters had somehow found the SSNs for inmates on death row, I think it was in Texas or Louisiana and had created synthetic identities around that real SSN. And so it looked legit, but you know, you, nobody knew. Right. So they would go through the credit washing, so they, they do a bust out, steal a bunch of money, then they go wash the credit, credit history and then they do it again with the exact same identity. Yeah, it was, it's pretty incredible. So let me ask again because, and we're going to talk about some strategies how to fight this in just a moment. I definitely want to make sure we spend a good bit of time there. But I know one of the other big challenges, specifically because it almost seems like there's some FCRA rules that you've, you know, you've got to work with the, within the confines of FCRA to figure out first party fraud. So just a quick definition of first party fraud for those who aren't familiar is it's when you've got a legitimate borrower who's applying for a loan but has no intent to pay that loan back, or they might be, you mentioned earlier, you know, where you're, it's maybe they're doing a straw borrower where they're borrowing on behalf of someone else or they're inflating their income or whatever and hoping that you won't catch it. How have you seen AI in particular used in first party fraud? Or is it being used to help individuals commit that kind of fraud?

Justin Davis: So I think it's very similar to what the CPN and credit washing is where we see it probably the most, as well as document manipulation. Just like on the pay stub, you'll see it primarily because it just makes it easier for them to supplement their lie with what looks to be a truth. And so say with first party, it's tough because when I think of fraud, I don't just think of intent, I think of the. In order to prove fraud in court, you have to prove three things. You have to prove material misrepresentation of fact, reliance on that misrepresentation by the victim and loss associated. So no matter if they intended to pay or they didn't intend to pay, if they lied to get that approval, you can consider that to be fraud. They did not tell you the truth, it changed your decision. If they would have told you the truth, it would have been different and you suffered a loss because of it. And so you will see, and you will see like honest people that will manipulate and change and doctor elements of an application or documents in order to. Many times they feel forced due to maybe economic constraints where they will lie and fabricate information in order to get approved for what they may have been approved for just a few years ago

Nathan George: or maybe to get a better rate. Maybe to get a better rate, to keep the monthly payment down lower where they afford it or something like that.

Justin Davis: Exactly. So they had the intent to pay, but they never actually could. They didn't have the ability to. And so maybe three, four months down the road, they can't afford the loan anymore, they can't afford the car or whatever they got the loan for and they end up defaulting and charging off. Well, if you would have known at the start that they were telling you a lie, you wouldn't have to suffer that charge off. So we would still consider that to be fraud. Now, in terms of AI for first party, it really is going to be the scalability of documentations and being able to manipulate documentation. Because if I am who I say I am and I'm just lying about where I work or how much I make, I need to doctor that information if you're asking for it. And so in order to do that I'm just going to literally use a pay stub generator and send it off.

Darren Thomas: I'll add one thing to that. So it really, it is kind of like a, it's an arms race, so to speak. Right. With what AI is doing from the fraud prevention side and what it's doing by the fraudster fraudsters. Right. And in the terms of first party, first party fraud, what's really helpful is when you're able to see that individual before and notice that that deviation of pattern has changed. Right. Has there been an abnormal pattern? And that's what we really try to do is through our consortium approach. Right. If we see that individual eight times, we have a pretty good track record for that individual. And being able to just detect those anomalies, extremely helpful.

Nathan George: That's just like how when the bureaus began to introduce the trended credit history, that greatly increased the predictive power of the models of the score models and things. And same thing with fraud that AI gives you the ability to take advantage of. I want to ask one of our other poll questions real quickly from the audience on this topic before we move into solution, kind of the solution phase. I want to ask everyone, it'll pop up here in just a moment. As you look at the fraud that you see in your daily business. What, what have you seen increase? What types of fraud are you seeing have grown the most in the past year? Yeah, again it's a, it's a little bit of a tricky question because some of these are done together of course, or in conjunction with one another. Interesting. Well, thanks for the output. It looks like synthetic fraud wins the battle here and first party fraud about growing, about half that rate, at least from what you guys are seeing. So thanks for those answers. Let's move into how AI, what works with respect to AI and using it to actually prevent this fraud. We've mentioned it a little bit, but Darren, I'll throw this question to you. Let me ask it much more specifically. If you're talking to a prospective client and they're asking you, hey, we're seeing an increase? Let's say it isn't synthetic fraud or first party fraud, one of the ones that are a little harder to detect with traditional methods, what's the top two or three things they should be looking at that AI can give them a real leg up to combating this?

Darren Thomas: Yeah, I'd say the first thing is really understanding the trend for that individual and relying on things beyond just the Social Security number. So what we see a lot of the times with synthetic and Justin can dive more deep into this is slight changes in the application. Right. So if there might be the SSNs changing, it might be first, you know, the name, occupation. So being able to tie that individual again to that kind of time series approach of like how often we see them, what's the change is I think a major key there. Justin, anything on the synthetic approach?

Justin Davis: Yeah, I was going to say on the synthetic side, a lot of it too will be around looking at the individual credit attributes for the consumer. Right. So does the credit actually make sense to the consumer and to the information that they're presenting? Right. Many times you'll see flags around, hey, the SSN is on the Death Master File. Fraudsters love to use SSNs issued to deceased people. Or maybe the SSN was issued before they were actually born. Right. And so they were born in the 80s, but the SSN for some reason was issued in the 70s. Well, it probably doesn't belong to them. But then those are more of the egregious errors. But the others will more so be, hey, you know, this SSN just started reporting, but this person's 50 years old, right? They, they should have established credit, but they just started reporting and then digging into, okay, well how was, how was that credit? You know, they have their score, everyone just goes off the score. But how is that score actually created? You know, looking at piggybacking where it'll be, I see five trade lines reporting to this person, but four of the five are authorized users. Right? And the only true primary trade that that person has is maybe a Capital One $500 starter card. That was one thing that I noticed when I was first starting to investigate synthetics back in like 2014, 2015 was I'd get a tranche of confirmed synthetics that we had looked at and investigated. And when I dug into their credit bureau, like why did each of these identities has a Capital One $500 starter card? Well, come to find out, it's a tri-bureau pull, or at least was at the time. I'm not sure if it still is. And so as soon as that was opened, all three bureaus were now the identity was now created at all three bureaus. And so whoever was the primary pull for a lender, they're going to see that person. Right? And so they're covering all of their tracks. So really not just taking the score for what it is, you know, looking at how was that score actually created. And then Nathan, to your point is synthetics will also wash their credit. Why go through the struggle of creating a bunch of new identities and having to go through that process of establishing a credit history when I could just bust out, wash it, bust out, wash it, bust out, wash it. Because when it comes to the bureau, once it's. Once it's off the record, it's off the record. You're not going to get that information anymore because it's cleaned.

Nathan George: And all their buddies on Telegram are telling him, hey, lender XYZ doesn't check for synthetic fraud.

Darren Thomas: Go here. Exactly, exactly, exactly.

Nathan George: So I want to change the use case a bit because you guys have a particular area of expertise in the auto space. Got a deep background there. And I think there's application not just in auto, but there's a lot of lenders like banks and credit unions with branch locations, or maybe you're doing point of sale retail finance and things like that where you've got people face to face applying for credit right in front of you. Talk to us a little bit about what you've seen in fraud when it comes to live interaction where you're applying for credit. What are the big differences in strategies for finding vulnerabilities as well as protecting from those vulnerabilities in that kind of hybrid model?

Justin Davis: So it was funny when I did a webinar back in 2017, I think with the Federal Reserve Bank of Boston, and then Ken Miser from ID Analytics, who's now at LexisNexis. And I always thought synthetics was primarily digital, right? They would never. Why would anybody ever go into a dealership and have to like get in front of a person when you could just do it from home? But Ken had told me, he's like, we're starting to see synthetics go into the dealership a lot. So once I started to come to. Once I came to Point Predictive, I started to really see it. Like, I could not believe it. My mind was blown. Now, there's a couple of different reasons why this is, is from after COVID, there was a ton of new fraudsters created because of the government unemployment. There were so many people that were there there, it wet their whistle where they never had committed fraud before. But it became so easy during COVID

Nathan George: PPP loans were real easy there.

Justin Davis: Exactly. And so once, once the funny money dried up from the government, they're like, okay, well I need to keep this train rolling. What's the next best thing? Next best thing is auto. A lot of the lending in auto is a little outdated and the ticket is much higher. Right. They can get, they can go and they can turn a car pretty quickly. And so people started to just go straight into the dealership. Now the other thing to keep in mind is the dealer is also an intermediary between the lender and the consumer.

Nathan George: Right?

Justin Davis: And so now you have new risks there. It's not just on the digital front where I'm going directly to my lender and I'm going to get a loan. It's like, now there's an intermediary in here and now I have to understand the risk of that intermediary could be

Nathan George: an inside job or it could just be they're not doing their homework when they're signing them up. Right.

Justin Davis: How is that salesperson at that dealership incentivized? They're incentivized to push metal and to sell cars. They don't care who they're selling it to. Especially if the lenders they're selling it to don't really monitor or they don't really have the need or the worry of, of having to buy back that car. Because maybe part of their, their contract with their lender is they're not going to get a pushback and it requires the lender to confirm fraud. And in many synthetic cases, they don't confirm it. So the dealership, a lot of the times can be perpetrating that fraud, whether they are colluding with, just not caring or even creating the synthetics. We've seen dealers more so on the independent side. Not many of the franchise fall into this category, but many independent dealers that maybe they don't even exist or they are creating synthetics to send off to their lenders just to push metal. And so that's, that's kind of what, what blew my mind when I started to really get into the auto space was it's not just, do I understand the risk this borrower presents to me, do I also understand the risk of the dealership that they're sending the loan from?

Nathan George: So just like in the, let's say a fraud strat or, you know, you're looking at KYC and you're looking at phone signals, you know you're going to flag. If you're doing it well, you're minimally going to be flagging, hey, this is a prepaid phone, right? These are burner phone. This is a burner phone network. The likelihood that it's fraud is going

Darren Thomas: to be high, but comes from here.

Nathan George: You're saying you need to do the same thing. You need to have similar checks and flags for specific types of distribution channels where your products are being sold, right? Like that's. Yeah.

Justin Davis: Because if you don't understand where the fraud is coming from, it's very hard to stop it.

Nathan George: Interesting. Darren, I think, did you have something you want to add?

Darren Thomas: I think Justin said that really well. I think it's just a reminder, right. The more data you can get from different sources, whether that's from the dealership, because fraud is particularly changing so fast. Being able to detect the anomalies from, hey, now there's a new trend showing up and it's from the dealer side or it's from, you know, now we're seeing a rise in synthetics. Just being able to catch that deviation in your data, even without a model in real time is super valuable.

Nathan George: So there's kind of shifting into the prevention portion conversation. There's a lot of vendors out there and there's a lot of people doing really good work, but it's really hard to know where do I start?

Justin Davis: Right.

Nathan George: Like, you know, the average risk officer or fraud officer, they're sitting there, they've got a lot on their plate. It takes a lot of time and resource and frankly, a high level of skill set to test new technologies and try them out without disrupting your current business. Can you talk a little bit about that? What, what, what have you seen has been effective in the testing and learning and implementing. Like what, what's been helpful for clients that you've seen to be successful to, to maybe ease some of that burden.

Darren Thomas: Every place needs to start is what data do I have and do I understand it and is it accurate? Right. Because any machine learning model or any fancy thing you throw at it is only good as the underlying data. So there's the data piece. Is your data clean. The second piece would be. Is it unique? Right. Am I missing things? Does it. Telling the complete picture? So when it comes to, you know, the model itself, you have to take those things into consideration. We work with a lot of institutions who have data science teams, big ones, and who don't. But at the end of the day, what we feel like, differentiate, differentiates us is our data.

Nathan George: Right.

Darren Thomas: So being able to leverage different data from a consortium, from lots of different lenders and individuals. Right. Is where I think you need to kind of start. Right.

Nathan George: So understanding the data, that's. That's important. I would assume going along with that would be having an infrastructure that helps you get access to that data and the tools to be able to test it alongside what you're doing at the same time. Those are all helpful things. If you had a prospect come to you and ask you, hey, you know, I've got a fraud prevention strategy in place. I know there's some things slipping through the cracks. What's the advice you would give? Or do you have some advice that you could provide and say, you know, here's the first three signals I think you should look at. Do you have these three? This, this, this gets you 80% of the way there, or like, what gives you the most bang for your buck when you're trying to figure out where do I spend my resources?

Justin Davis: A couple of questions I typically will ask is first, how are you defining fraud? Because they could say, I have this fraud prevention program, fraud prevention strategies. Things are slipping through the crack. Okay, well, what is fraud to you? Right? Because many lenders I'll talk to, they're just, it's, it's a claim of identity theft. I got someone who told me that they're a victim and I've tagged that as fraud. And that's what I'm looking at. Synthetics, maybe. But how many are you truly catching and are you are able to properly identify them and categorize them? And so it's, how do you define fraud and then understanding what are you doing today and to, to prevent it and then how do you know that things are slipping through the cracks? Right? So it's like if it's. I have the fraud prevention flywheel is what I typically will kind of train and teach is I have, you have a prevention stage, you have a detection stage, and you have an investigation stage. And they just feed each other. Here I'm preventing fraud, detecting how, how good am I at detecting the stuff that, that, that got through? How good am I at investigating the root cause of that? And then can I feed that back into my prevention strategies? So what are you doing in that detection stage? And then I think practically being able to kind of broaden your horizons and understand how much risk that you're probably looking at, how much fraud you're actually experiencing is looking at not just your first payment defaults or those that are tagged as fraud, but digging into your early payment default, the first six payments, if they stop paying that for six months, that is a very, very strong proxy for fraud for both third and first party.

Nathan George: Right?

Justin Davis: Like to my example earlier is they may, they may have lied to you about how much money they made and they had the intent of paying you, but three months, four months in, they stop, we would still categorize that as fraud. You should have known that they were telling you a lie about how much money they made, they could never afford that purchase. And so look at your early payment defaults, those within those first six months and start to investigate that. And I think you'll be very surprised at how much fraud is actually hidden in your credit losses.

Nathan George: That also sounds like open banking data or having some way to calculate affordability is really vital as well. Like actually looking at that monthly cash flow could be a real helpful one.

Justin Davis: Right? Not just, hey, the credit score looks good and they say they can pay me, right? And now I have a fake pay stub that says, yes, they can pay me. So I'm going to go ahead and, you know, book that loan. It's like that doesn't work anymore.

Nathan George: You don't want them going Netflix or paying me back. Yeah. So let me ask a couple more questions. One that is interesting to me is we've talked a lot about, you know, the benefits of AI to us fighting fraud, the benefits of AI to fraudsters committing fraud. What are some risks in using AI to combat fraud? As you look at what you've seen clients do or attempt to do, are there any big risks or landmines that you could help us with? Say, hey, don't step on this one.

Darren Thomas: So I'd say at the end of the day, AI is powerful, extremely powerful. It's not magic. You know, the, I'll go back to what I was saying earlier. It's only good as the data you're feeding it, right? So you need to make sure that information stays, you know, complete and not outdated. You could end up missing a lot if you, if, if you start to change your own data behavior. Right. And I would say too, it's just the, it's, it's not a plug and play solution, so to speak. It's a living, breathing part of your risk strategy that needs to be fed monitoring, kept in check. And how you can do that, right, is not getting complacent with the current model. So something like we do is we refresh every six months because we know fraud trends change and we need to make new features as a part of that. Right? So it's never being complacent in the sense of, hey, I have, my model works well today, it'll work well next year or next month as well. Right. So it's that evolution, I'd say.

Nathan George: So if I, you know, if I'm sitting here thinking, and I might be getting out over my skis on, on this question, but I mean, what the average pace of change for a credit risk model, right. You're, I would think your fraud model needs to be updated, optimized, whatever, at least that frequently and probably significantly more so because Telegram probably funds, exploits and distributes them a lot faster than the credit risk side.

Darren Thomas: Right.

Nathan George: I mean, would you agree with that or comment on that?

Darren Thomas: I would totally agree with that. And I think it's one of those things where it's good to have the six month rule at the minimum. Right. But you can start to see shifts if you're watching the data that's feeding your model. I keep going back to that. Right. And you're seeing changes in that data that start to spike in a certain week and it's different. Right. There's lots of ways to look for data deviation. You can catch that before the defaults start to add up. It's tracking and monitoring all the things that you are seeing in an efficient way through advanced data science techniques. So that's how you stay ahead. So if you see that deviation, we see it all the time. We need to make a new feature to detect this if we're not capturing it as well as we could be.

Nathan George: Well, that takes me back to what you said earlier, Justin, as well. You mentioned the importance of defining what fraud is and how you're measuring it because it's always easy to get lost in the, oh well, this was a market shift, you know, losses, our loss ratio changed because the markets changed. But you know, if you don't have that fraud definition, you're always trying to figure out what's what. But if you've got the fraud definition, you've got a stake in the ground that you can compare to and see how you're doing.

Justin Davis: I like to say fraud is preventable where credit risk is a lot harder to prevent. Right. It's like if someone loses their job, there's, there's nothing I can do to prevent that. But there's, there's red flags that are present on applications to prevent fraud. If you know what you're looking for, you can stop it. If you have no idea what you're looking for, you're not going to stop it. So it's very important that you have, you have a proper way of identifying what fraud is. Because once you've defined the problem now you can address the problem.

Nathan George: So one more question, maybe two if we have time and then we'll open it up to the final Q and A. When you're thinking about first party fraud in particular and how do you distinguish between or does it even matter? I think you kind of already answered this one, Justin. But what's the best way to spot first party fraud ahead of time? And is it important to differentiate, you know, which of these was intentional versus not intentional? And in my head I'm connecting that to treatment strategy on the other side, like how do I deal with this once I, once it's observed, once it's happened, you know, is this somebody I want to try to, you know, negotiate with or change the payment schedule somehow or whatever the repayment schedule. Could you speak to that a little bit? And how does, how does that fraud calculation play into that?

Justin Davis: It's a good question because it gets down to intent. But to my earlier definition, if they're lying to you and you're going to suffer a loss for it, we would consider it to be fraud. Do you want to book that loan? Now? It is up to, it's the risk appetite of the lender. Right. It's going to be what are they comfortable with? But it doesn't have to necessarily just be, I'm going to decline this application. It could be if I understand and I'm seeing that there's maybe some first party indicators here rather than just third party, maybe there's better offers, lower offers, maybe, maybe you can counter offer in a different way, offer them different terms, different amounts, different payments. And if they have good intentions, then they're probably going to want to work with you on that. If they don't have good intentions, they're just going to walk away. But it could be outright, just decline it. It's very likely that I'm going to suffer a loss because of this. I don't even want to deal with it. I'm going to have to have a team that underwrites it. I'm probably going to have a team that collects on it. It's just an operational nightmare, right. Or maybe I want to give this person a chance because maybe they are just in a financial struggle and maybe I'm just going to offer them less and make it more easy for them to pay the loan back rather than just try to get them into, I'm going to get them into a vehicle, I'm going to get to, I'm going to get them a credit card and I'm going to offer them a personal loan.

Nathan George: Right. So it would be fair to say, depending on your risk tolerance and your goals and your growth goals that you've got, you know, it sounds like connecting your fraud strategy and the type of fraud to your credit policy could be, and the offers that you're putting in front of them and the terms and those things could be really useful to have a little more complexity in that to hit your goals.

Justin Davis: Definitely. Like one example would be income misrepresentation. So someone lying about how much money they make. We actually see that if someone is misrepresenting their income, it's actually negatively correlated with default. So the default risk is not strictly tied to how much money someone makes. But once you start to add on additional risk elements, the default exponentially grows. So it's what's the severity of the lie here is also a good indicator to understand what the intent is of the person and how you can treat them.

Nathan George: Interesting. Or are they just bad at math? If they're paid hourly, like how many hours in a year again?

Justin Davis: And who's it coming from?

Nathan George: Right.

Justin Davis: Maybe it's not the borrower. Maybe the dealer that's the F&I manager is just plugging away numbers.

Nathan George: Right? Interesting. Yeah, because they want to get them in that car. Another thing that was curious or interesting out of your report was the type of income made a difference to the fraud signal. Or at least it made a difference to income inflation. Right. Like could you maybe do a quick blurb on that? Because I remember reading about gig workers and self employed in particular. There were some differences between what you saw with the income numbers.

Justin Davis: There's a couple of different risk triggers that we see that we're monitoring because when they're present, the default rates kind of skyrocket would be the people that hide their gig work. So we see them across our network where they're saying they drive for Uber or for Lyft or for Uber Eats or whatever. But they're telling the lender that they are a W-2 earner and they work somewhere completely different. They're not disclosing that they do this other work. Do they actually work at that place? Or are they just gig or they're hiding, they're hiding their retirement or maybe disability where they were retired or disabled, but now they're not. Right. And so they're fabricating their employment. And then from a self employed standpoint, we'll see that too is they'll tell one lender they're self employed, they'll get declined and now just magically they become a W-2 worker.

Nathan George: Right.

Justin Davis: So they're just hiding that they are self employed and they're just saying, yeah, I work at Amazon as a driver.

Nathan George: That's another interesting one, is not just the amount of income and verifying it, but actually where's that income coming from? Source can be help helpful signal. One other question I wanted to ask while we're waiting on audience questions. Was there, maybe you can help with this when it comes to real time data and historical data, like historical data and it has to be live data right now. It's interesting because a lot of times you're trying to test something new, you're doing a retro with data that's old. So how do you deal with that conundrum when you're testing new fraud solution? And what's more important, is historical data important at all outside of training? Like how do you balance that?

Darren Thomas: I would argue they're both, they're both important. And I'll start with the historical side. So like you mentioned on training, I think that can't be overstated how important training a good model is and making sure you have good data that you're training on and you're able to show results in the past. Right. But how do you translate that to real time, real time data? So I think how you can track it in real time data and as, as you're scoring in production, right. Is are those deviations that are changing catching the things that you want to catch? Are you seeing the, the first indicators of, hey, we know this is synthetic. We found it a week later. Did it, did it catch it? Right. So real time can really be used as your, your measuring stick to see if you're at the right place. So historical really to train, real time to detect, make sure it's working and you know, see if there's any anomalies that need to be then captured.

Nathan George: Yeah, that's interesting. I was having a conversation with, with a company that specializes in email fraud signals based on mostly email data. And one of the things that they shared was that it's very common for them to see email used one time for fraud and then you'll never see that email address again. So looking at age velocity, you know, when it first appeared that kind of thing was really interesting. But as a consortium flag or as a fraud flag, flagging that email in the future had little to no value in their estimation. That was an interesting one that I hadn't thought about before. It's funny though, I have seen some of the fraud dashboards you'll use and it's just hundreds of applications all fraudulent from the same exact email. Maybe they tumble the, the front part of it a little bit.

Darren Thomas: But yeah,

Nathan George: one other, one other question that just came in. When you're testing and rolling out new fraud tools, are there any technical challenges, operational or process challenges that you could put in that you can overcome or plan ahead to, to avoid. Like what, what are the things that you've seen have been the most biggest obstacles, biggest obstructions to testing your solution when you have prospects come to you and want to try it out.

Justin Davis: So for, I'll kind of hit it in two different ways, testing and implementation. So for testing, right, it's I think it's understanding the data that you want to test and understanding the problem that you want to solve and knowing what, like how are you going to measure success? So sometimes it's just, let's just throw data over the wall and see what comes back. Like, yeah, but you're not really going to know what you're, how are you going to measure this? How are you going to analyze it? Like what are you actually trying to do here and accomplish. And so being able to do the work up front to know what you were really trying to solve makes the testing process much easier and it's the decision making process much easier too because now you have more of a, a rigid decision making. Did it, did it solve what I was hoping it would solve or did it not right. But from an implementation standpoint, I think first question is have they done it before? Right. Has this vendor actually done this before and if so, how many times, who with? Right. You can do reference calls, but what I've seen and what has been very successful for us is having, if we've done retros and data studies, we'll build out recommended strategies and kind of like a phased approach to say, hey, here's how you're going to get the most out of the solution going forward from day one onward. And how can we kind of get you comfortable? Because especially when a lender is new to AI machine learning, whatever you want to say, they're a little hesitant, right, to kind of open the floodgates and allow it to do what it to trust the score, as we like to say. And so we'll phase it in like, hey, just start on the tails. Like be very conservative with how you're going to use this before starting to get a little bit more aggressive as you understand how it works and you start to trust it more. So being able to have kind of a best practice into implementation as well as hey, even if we've done a retro, here's custom strategies you can use makes it a lot easier rather than we're going to implement it, we're going to turn it on and then you got to figure it out yourself.

Nathan George: Sounds like a great question for a prospective vendor that you're evaluating, right? Do you guys have a not just tell me how great your fraud tool is? But hey, how about how would I implement this? How does it work for me that I can follow? Yeah, yeah. That doesn't, you know, completely turn my numbers upside down on the origination process.

Darren Thomas: Yeah.

Nathan George: Well, we've got one minute left, so I want to throw a big thanks to Darren and Justin to both of you for bringing the smarts today. Appreciate everyone's questions and insights that were shared. Before we log off though, in our last minute, I want to tell you about two resources that are going to be available. They'll be sent out to you via email. There will be a recording of this made so you'll be able to access that later. There's a fraud and credit risk eBook from GDS Link that's available. It's a great resource that talks all about both credit and fraud risk and how to test, learn and execute quickly and well. We also have a copy of Point Predictive's Fraud Trends Report. We've mentioned a lot of examples out of it. It's a great resource for you. It's full of some real examples as well as some great solutions for fighting fraud. You'll get all of these in your inbox, in your inbox shortly. Thanks again to everyone for joining us. And remember that the fraudsters might be fast, but you only have to be faster than your slowest competitor. So with that, thank you very much and we'll catch you next time.

About Point Predictive

Point Predictive helps lenders make smarter, faster decisions by making fraud more predictable. With a proprietary data repository built from millions of credit applications, they offer unmatched insight into patterns of risk. Their patented Artificial + Natural Intelligence™ approach blends machine learning with expert intuition, enabling financial institutions to detect fraud earlier, reduce losses, and create a better experience for borrowers. Learn more here.

 

About GDS Link

GDS Link makes modern lending simple. Our real-time decisioning platform integrates over 200 data sources with advanced analytics, enabling lenders to make fast, data-driven credit decisions while mitigating risk. From loan origination to collections, GDS Link provides seamless automation, policy monitoring, and AI-powered insights to help financial institutions optimize lending strategies, stay agile in a competitive market, and drive better outcomes. Follow GDS Link on LinkedIn here.

 

Resources Mentioned:

Download GDS Link’s The Real Risk Equation eBook

Get the full 2025 Auto Lending Fraud Report from Point Predictive

Looking for more tools to reduce fraud risk and improve decisioning? Visit our Fraud Prevention webpage to learn more.