Masters of Risk | Season 3 Ep. 2: AI or Not: How Do We know What is Real or Not

View Full Transcript

Stewart Webster   00:00:46

Welcome to Masters of Risk Season 3, Episode 2. Anatoly Kvitnitsky is the Founder of AI or Not a company that helps individuals, businesses and institutions determine whether content has been generated by artificial intelligence. With a background in identity verification, fintech and regulatory technology, Anatoly brings a sharp risk management lens to one of the most pressing challenges of our time, distinguishing human truth from machine-made content. Before founding AI or not, Anatoli helped scale innovative technology companies at the intersection of compliance, trust and data. Today, he's at the forefront of the generative AI frontier, building tools to foster trust in a world where seeing is no longer believing. Anatoly, thank you for joining us today. 
 

Anatoly Kvitnitsky   00:01:39

Thanks for having me, Stu. Incredible to listen to a speech about myself from one of my fraternity brothers. We've come a long way since New Brunswick, New Jersey. 
 

Stewart Webster   00:01:49

Absolutely. And glad that you agreed to join us on the podcast. Now as you mentioned, I know a lot about yourself and about the company that you founded AI or not, but many of our audience members may not know about AI or not. So I guess let's just start with the basics. Can you give a quick overview of the company, AI or not? And what does the company do? Who are your customers? And what problem are you solving? 
 

Anatoly Kvitnitsky   00:02:15

Yes, happy to. So I started AI or not with actually one of my dear friends, Johnny Nelson, who I've known since second grade about 1.5 years ago. And the reason I started it was when I saw what was happening with the innovation in generative AI, I did not think there was enough being done to try to fight the dark side that was coming from it. So like you described, I spent a decade plus across financial services, risk, compliance companies, big and small. And every -- throughout that decade, I've seen new technologies get used for bad every single time, whether it was social media and the creation of misinformation that came with the post on it, whether it was cryptocurrency in the way that it was used for a new store of money, but at the same time, a new way to launder money. Generative AI is no different. In the hands of someone trying to do good, it will do a lot of amazing things. make us more productive as individuals, help us a lot more. But at the same time, in the hands of fraudsters and bad actors, it will have the same effect, make them more productive and make -- help them do the bad things that they want to do. So that really prompted me to start AI or not. And what we do is we've built a suite of AI detection tools, starting with image and then we went into audio, video and text to help people determine and companies determine whether a piece of content is real, human-generated or artificial intelligence generated. And that's what we've been working on, and that's what the company does. 
 

Stewart Webster   00:03:51

Well, that's fascinating. And definitely interested in learning more in the conversation here. So when I think about AI, right, that's the whole buzzword nowadays. For the past year or so, it's been artificial intelligence, AI, large language models. When I started my career 15 years ago, the whole buzzword was big data, right? Now the buzzword is AI. So I'm just curious, it seems like you launched your company during the time of massive AI hype and rapid evolution. So I'm just curious, what are some of the biggest risks and uncertainties you face during the early times. When I think about technology companies, especially ones that are private equity-backed or venture capital type companies, there's this whole concept of the J-curve, right? So the early years, there are some losses. And then after that, there's rapid growth. And if you were to plot that phenomenon, it would form a J. So just curious on your side, what were some of the uncertainties you faced? 
 

Anatoly Kvitnitsky   00:04:47

Yes, all really good points. I think that the new advent of generative AI will make -- like makes almost big data look like small data now, to be honest, because these generative AI companies like to build the LLMs that they are, you're ingesting all the data of the entire Internet. Big data, you're right, it used to be just the company's data, and now that's microscopic compared to trying to ingest the entire Internet. And that's what companies like NVIDIA is powering, cloud providers are powering all for the invention of these AI tools. As far as the J-curve, it's a great theory and practice. Before this, I used to be a venture capitalist for a number of years, and it's 100% accurate. There is a -- like you said, in the beginning, losses and the invention of sometimes new categories like which is the one that AI not plays in. 18 months ago, 2 years ago, this category didn't exist AI detection because there was not enough AI that could fool anyone. And now it's -- now we're riding along that J curve, just like you said to do. And there's a period where companies have to build out their product and some products are more costly than others. And I think AI outside of hardware is up there with like the amount of cost and the amount of effort it takes to build, but it's also what builds moats. So AI in a lot of ways is actually what creates moats and creates sustainable products as opposed to just software, which AI can actually very much build themselves now. 
 

Stewart Webster   00:06:30

Absolutely. So just curious if you could talk a little bit about the risk of AI, right? So I heard a lot of things about hallucinations, biases, deep fakes, I'm not too sure what those are. And I don't know if our audience knows. So I guess if you can kind of expound on what exactly those are. It seems like AI has really advanced over the past couple of years. If I were to look at my social media account last year, I could tell what was AI and what was not AI simply by the facial movements or how the lit movements didn't match the actual speech. But now it seems almost impossible to tell the difference, right? So I guess, can you break down what's at stake? And then also, what are the risks out there, like I mentioned, hallucinations? 
 

Anatoly Kvitnitsky   00:07:16

Yes. If I can break this down almost like in quadrants on a per modality basis and for people and for companies. So let's start with text. This is, I think, the most commonly used. I think publicly like ChatGPT announced they have 0.5 billion users already. It's probably -- that information is probably dated in the couple of weeks that they posted. And that's primarily text-based. And the risk there that we run against is if we are using this tool as a trust mechanism for answers for information, it does spit out a lot of -- still a lot of false information. They've been trained to really try to please the end user to really make them happy with the response to the point that they actually make things up to make you happy. Like if you're asking for a piece of information and that information actually doesn't exist, it will do its absolute best to make it exist and actually even produce fake URLs, which I've seen it produce. So -- and that -- we've seen this outcome for lawyers citing not real cases that they actually ChatGPT for professors creating false content that they're teaching against. And I think even for people making decisions in their day-to-day lives, if you're trying to gather information about politics or something you're concerned about and you're trusting an LLM and it's indeed spinning out things like thinks you want to hear as opposed to actual fact, I think that's a huge risk. On the company side, too, there's a lot of, call it, shadow use of generative AI tools. So a lot of these things are blocked within corporate environments because it's unclear like what gets used for training, what gets not. So a lot of times, it's shadow use. Maybe it's bring your own device or maybe it's just within -- outside of the company walls, but still very much used for work. And there, especially when you get to tools like DeepSeek is a really great example. And everything gets stored and you can start tying back maybe not public information or maybe things that are very proprietary to companies and all of that gets stored some way, some shape or form and possibly even used for training. And whether or not that information will ever get spent back at another user with the things that you've said is very -- is a huge, huge risk, huge risk if you're using cloud-based AI for a company as opposed to on-prem and like owning your own stack, then a lot of that risk goes away. On the image side, Stuart, it's exactly like you said, the days of counting fingers, the days of like perfectly smooth skin generated by AI, none of that applies anymore. All these people are very lifelike. They are perfect in every way, shape or form and sometimes imperfect to make you believe that they're actually real. So what that looks like from a consumer perspective, it's -- you're right, the online experience is absolutely flooded with these, and you can't tell anymore. Like there's quotes like, "Oh, I think I've been seeing less AI-generated content online. And you're like, Oh, wait. There's actually studies that project in the next 2 years, 90% of all content will be AI generated, which to me would be a very sad state of the Internet because I don't think that's what we go on forums or social media to look at. And then from a company perspective, if you can create realistic images of people, of documents, of receipts of bank loan application or whatever the thing is, all the processes that rely on human vision I think, are at huge risk. So whether it's analysts checking KYC documents or whether it's risk analysts looking at all the documentation provided to make a lending decision or receipts even for expense reports. If all of that can be doctored with AI, all of those things are at great risk. On the audio side, I think there is a lot at stake with recreating people's voices, whether it's a musical artist and you're taking away from their brand, from their way to make money, you're impersonating politicians. This happened where there was a fake Joe Biden phone call that called an entire state, New Hampshire, I believe, to tell people not to vote, and that person went to jail for the record once they found them 3 months later. But there's a lot to be -- there's a lot of risks when you can recreate someone's voice with roughly a 15- to 30-second snippet and then have them say anything. And then the tech side to me, I think, is actually the least risky. But then when you start combining modalities like image, audio and then video that combines all of that, it's really hard to discern what's really what's not anymore. 
 

Stewart Webster   00:12:26

Well, definitely a lot to take in there. Just curious behind the technology that AR not uses -- AI not uses rather. So what type of technology do you guys use? Is it a proprietary model that you guys have or proprietary code that helps you identify which text is AI-driven, which audio is AI-driven? And just curious for the average person, how can we kind of distinguish between what's AI or not? 
 

Anatoly Kvitnitsky   00:12:52

Yes. All good questions. So I'll cover the first one first. So we built all of our AI detection algorithms in-house. To me, it was incredibly important to do so just to make sure I'm able to tell all my users, whether they're the 300,000-plus free users or the hundreds of companies and even government who uses us, all of the data that you input gets deleted immediately after a check. And if I didn't own all of that myself, if I didn't build all of that myself, I could never make that statement. But it was really important for me from the jump. For images, what we do is we have vision models that are trained on pixel detection patterns for each one of the generators. Each some of the generators like whether it's the Google One that was just released or Midjourney or Grok, they each have their own distinct pixel pattern and all of those are different than, say, what comes from an iPhone phone or a canon DSLR. And we have trained computer vision models to detect and to discern the patterns between the 2 types. For audio, we're analyzing wavelengths. -- of what constitutes like a real person's voice versus real music, and we have models for both of them. And there are some pitches and some sounds that may sound identical to the human ear. They are not to a computer and the wavelengths are different enough that we're able to tell with a really high confidence rate. On the tech side, we've built transformer models that can tell whether certain words or phrases are used more commonly by AI or not. So before -- and so the way you think about all written forms is November 2022 is when ChatGPT launched. So there's a strong case to be made that before that date, all text written is human generated. After that, it's kind of up in the air. And there's definitely some words that start coming up a lot more frequently post November 2022, things like Delve, things like dashes and overuse of dashes, things and certain phrases that I'm sure we've seen that we can be like, oh, that's AI. So the system really takes apart those words, phrases and sentence structures to be able to determine it. And yes, that's how it works. How can people be able to tell today if the visual -- like you said, Stu, if the visual cues start looking so real, you almost have to look for the context within the content. Like it's -- I hate to say it, but nowadays, it's almost like you have to question everything because sometimes people even accuse real things of being AI, and this happened recently in a French protest and Turkey -- and people are accusing things of being AI when sometimes, in fact, they're real. And then sometimes, in fact, they are actually AI and they're being reported on as real. So it cuts both ways. So sometimes it's almost the context that has to be thought about too when you see a piece of content. And on the written side, I think certain words and certain structures are definitely skewed towards AI. And when you see things like that, and again, the context feels and looks off, I think it's a really great red flag to determine that it's actually not real. 
 

Stewart Webster   00:16:21

I definitely just learned a lot there. So I'll make sure going forward, when I e-mail my clients, I won't use Hyphens, and I won't use the word Dove. This way, they won't think it's AI generated. That's right. So when I think about our audience, many of them may be Board members or risk officers or really just the management level of corporations who want to start integrating third-party AI tools into their workflows, right? So I'm just curious what types of questions should those Board members or C-level executives ask of the third-party risk tools just to make sure that they're credible. What type of questions should they ask when they're evaluating those tools? 
 

Anatoly Kvitnitsky   00:16:59

Yes, it's a great question. The way that I would think about it is the data that you want analyzed or used by those tools, like how much -- how risky is it if someone else got their hands on it? And I think the decision then becomes, do you stand up your own solutions? Or do you use cloud-based API-based solutions? Like, for example, if you're -- if it's, say, medical records or financial statements or financial records that are not yet public, there's a strong case to be made to stand up your own. And though that sounds quite overwhelming, it's actually not that difficult to do. There's incredible open source models. even like the DeepSeek one, if you stand it up yourself, you own all the data in and out of it and there's a strong while you're still getting like 97%, 98% of like what the closed proprietary tools provide in the cloud. But if it's more generic use or analytics or you just want like the best possible model, then I think the API cloud route is a really, really great way and a quick way to get stood up really, really fast. If there's no proprietary data or data that can be tied back to any person, any PII or any individual or company that shouldn't be a very, very strong case to be made to use the cloud-based API ones. So that's how I would think about it about the risk reward for each. But at the end of the day, I can definitely see a world where many large enterprises like S&P is a really great example of having their own internal like in-house versions of these LLMs stood up on their own cloud. So they know where the data gets used, how the data gets stored while still creating and getting the benefit of creating your own AI-powered products. 
 

Stewart Webster   00:18:57

So it seems like AI or not, you're not only providing a product, but also a service as an adviser to -- and also as a consultant in terms of artificial intelligence and how companies should integrate that into their platform. So that's really cool to see. You were quoted before saying, we're entering a phase where people need tools to believe what they see. So just curious, what does trust look like in the world flooded with AI-generated content? Is it a blue check? Is it some type of certification? How do we trust? And how do we know whether something is human made? 
 

Anatoly Kvitnitsky   00:19:32

Yes. I think a lot of people are still basing their assumptions of they can see what's real or hear what's real or not based on things that are very much updated. -- like the finger thing, all the AI tools have solved. The text thing, they have all solved these things. The audio is perfect and crystal clear without any weird glitches. So there, it really becomes a matter of depending on the context that there's a lot of money on the line, like it's worth running -- it's worth doing a check. For large financial institutions and large financial transactions, I actually think in real life becomes a really important aspect as well. I think one of the huge beneficiaries of this will actually be banks in a lot of ways because they're one of the last places where we all have -- still have to go in person a lot of times to get verified, to get checked, and they have an incredible opportunity to be like we've seen this person in real life. We didn't just like onboard them with a selfie check and ID. And I think there becomes a lot of value like before. I think the days of -- there was a famous example in Hong Kong, a fraudulent $25 million wire was done over Zoom. I think those things are -- those days are over of doing transactions like that without a handshake or some kind of safe word almost like to make sure that this is really happening before someone who you think resembles the CEO of the company tells you to wire some money. So I do think we will have more checks in place, almost like a filter for is this actually AI or not? And then for really large, really high stakes things, I think there is going to be a huge benefit and a huge value add for the in-person aspect as well. 
 

Stewart Webster   00:21:37

No, absolutely. So where do you draw the line between empowering transparency and policing content? So what I'm thinking behind this is if you empower transparency, right, you allow innovation to spur and you allow technology to advance, right? And that's how we grow, and that's how that J curve, if you will, really takes off as a society. But then if you police content and you have a lot of regulation, you could see the growth of that technology. So I guess, how do you draw the line between the 2? Or how do you balance between the Wild, Wild West and where we dictate a highly regulatory environment for AI? 
 

Anatoly Kvitnitsky   00:22:16

Yes. It's a great question. And I think that question is playing out even like copyright laws that are being written right now in the United States and in the U.K., where every country is like no one wants to fall behind on AI innovation. But how do you balance copyright laws and how you balance like the training laws and what's allowed to be trained on and what's not. And then from the consumer side or we're all consumers at the end of the day, like do we want AI-generated content like flooding our feet, just like you said in the beginning, so which is like a really great way to start this episode. And frankly, if not all, most of the social networks are creating the AI tools to help you generate the content. So at some point, when people realize like most of what I'm seeing, like that influencer holding up that beauty product or this person in the gym or like this weird street interview that we see every fourth post is actually not real. People may start getting tired of that world. Like I would not want to go to any social network if like the projections are true, where 90% of the content is AI generated. Like I still want to know what's actually happening for real I still want to know that the people in front of me are actually real individuals with real opinions. I get it like sometimes I generate content, could be funny, it could be entertaining. But I would like to see the transparency and the ability to filter it all out. We're actually working on this on AI or not to help people do so within some of the plug-ins we're working on. But I do think on the copyright laws, I do think some loosening is not the worst thing for the fight of innovation while making sure like the creators that are being used get compensated in some way, shape or form. While on the consumer side, I do think filter systems and at least transparency around what they're seeing without being able to tell themselves is shown as to like, hey, this may include some AI-generated content. So I think transparency, but not stopping this like runaway train that is innovation and AI either. 
 

Stewart Webster   00:24:41

Absolutely. And lastly, more of like a philosophical angle. So I know there's this whole concept of the singularity, right? And from my understanding, that's the point where AI will kind of take over humans. So from your perspective, do you think we'll actually reach singularity -- and if so, what's the risk that keeps you up at night about it? And what's the time frame for singularity in your eyes? 
 

Anatoly Kvitnitsky   00:25:05

As a user of these tools and as an implementer for certain parts of my business, I think in certain areas, it's already here. So I think even when I'm building my company, I'm planning out like who am I hiring, I think we're down to like bare bones basics of what you need, meaning you need people building the thing you're selling and you need people selling the thing that you're building. All the in between of administrative analysts, someone crunching numbers, I don't think is as needed anymore. I hate to say it. I used to be one for a lot of my career, but my first job at ChatGPT can do a much better job than I ever could do it and much, much faster as well in areas like even development. My CTO estimates that for some pieces of our code like on the front end, which is like how the website looks and feels, 40% power is being -- that code is AI generated. The things that AI still can't do is actually build AI itself. So we still have to have people to build it. The personal aspects of doing business, relationship, sales, I think that becomes even more important because that doesn't go away. But all the things in between. I think if we're not there, and actually we're already seeing signs that I think we are, I think it continues to get better. I think the last step, and maybe AI will be able to do maybe not, is building AI itself. Then I don't think there's -- then I think things get a little scary if AI can start building its own AI and starts determining like what safeguards need to be in place, what protections need to be in place and it starts analyzing things like how do I save the planet and it comes to conclusions like who are the biggest -- who pollutes the planet. There's only one species that does so. And it starts coming to conclusions like that, I think we go into a really scary place. So I think that's the last mile that I have not seen AI be able to do, no idea when it can. But for a lot of that are happening today, I actually think we're already there. 
 

Stewart Webster   00:27:25

Wow, definitely a dynamic conversation. Anatoli, thanks so much for joining us on Masters of Risk. From start-up risk to enterprise threats and the future of trust itself, it's clear we're just beginning to understand the true impact of AI, not just on business, but on society. For anyone who wants to learn more about Anatoly's work, they can visit aiornot.com. And to our listeners, if you enjoy this episode, be sure to follow, rate and share Masters of Risk. I'm Stuart Webster, and we'll see you next time.