← The Digital Executive
The Digital Executive
Anatoly Kvitnitsky on Battling Deepfakes and AI Fraud in the Digital Age | Ep 1049
00:13:02
Loading summary
Transcript16 lines
- [00:00]
A
Foreign.
- [00:09]
B
Welcome to Coruscant Technologies, home of the Digital Executive Podcast. Welcome to the Digital Executive. Today's guest is Anatali Kvitnitsky. Anatali Kvitnitsky is the founder and CEO of AI or not, an AI detection startup focused on helping organizations safeguard against the growing risk associated with generative AI and including deepfakes, fraud and misinformation. Under his leadership, the platform offers tools that enhance compliance and protect the integrity of digital ecosystems. Prior to founding, Aryan Anatoly was a principal at American Express Ventures, where he led strategic investments in emerging technologies across commerce, payments fraud prevention, data analytics and cybersecurity. He also served as Vice President of Growth at Trulio, a global identity verification company where he oversaw strategic sales, partnerships, R and D, M and A, and fundraising, helping scale its solutions to promote financial inclusion and best practices in data privacy. Well, good afternoon Tully. Welcome to the show.
- [01:06]
A
Thank you for having me.
- [01:08]
B
Absolutely. I appreciate it. Making the time you're calling out of the Bay Area today up in Northern California. And I appreciate that I'm in Kansas City, so two error difference, not a big deal today. But Tully, I'm going to jump right into your first question. What inspired you to launch AI or not? And how is your platform tackling the rising threat of generative AI misuse like deepfakes and synthetic media?
- [01:31]
A
Yeah, really great question. I spent the majority of my career tackling fraud and KYC problems. First at the largest credit bureau in the world, actually down in Orange county, where you and I both resided at some point, Brian. And then I had a startup that became a unicorn in the KYC space. So in my career of a decade or so of fighting different versions of fraud in kyc, it was really a game of whack a mole of fighting against like new technologies and what they brought to the table. So when I saw fast forward to end of 2022, beginning of 2023, when I saw what was happening with generative AI then, even though at that point, you know, fingers were wonky, you know, the content wasn't exactly photorealistic, I knew it was only a matter of time before it was going to be. And that really inspired me to start AI or not, because I knew a lot was being invested into generative AI, but I didn't think enough was being invested into stopping the dark side of the technology. So that's where I really started AI or not for towards the latter half of 2023.
- [02:33]
B
That's awesome. And yes, we've seen such a lot of things happen in the Last just couple of years the way the technology in AI is leapfrogging. But I love that you shared a passion of yours. You obviously working at KYC fighting fraud and abuse but seeing that the technology capabilities is only going to get better that you jumped ahead and said you know what, there's a problem here or a potential problem going to be huge. Inspired you to start your company, AR or not. And I really love the story. Tolle AI or not helps organizations detect AI generated content. What are the key signals or patterns that your system uses to identify fakes?
- [03:10]
A
Yeah, it varies on a modality basis and we cover all. So we do image, video, deepfakes, audio and soon to be text. So on the I can go one by one and I'll and I'll keep it brief and then finally fire away with any follow up questions. So on the image side we there are a few different foundational ways that models generate images. Whether it's mid journey now ChatGPT4O or Flux, they all have distinct patterns in each one of their images that our machines can read and identify as AI generated. They have gone so good like the naked eye can no longer tell. Many times I look at this stuff all day, I can no longer tell. Elon Musk, one of the creators of Image Generator, among many other things, said he can't even tell what's AI or not nowadays. And that's an exact quote. And it really boils down to the patterns that each one of these generators make for audio. It's the wavelengths that the generators make versus what a real drum sounds like or a real voice, whether singing or speaking sounds like. Though the pitch and the style might sound exactly the same to the human ear. The computer and the algorithms that we create pick up the distinct differences in the wavelengths that each of them generate. For video. It's kind of a combination of all of the above. You're analyzing the frames, you're analyzing the pixels within those frames, frames of the video. And we also run through all the audio and then finally text. There are definitely certain words, phrases and combination of words that generators use much more frequently than that of people. And we pick up those signals as well. So it's really the overlying theme is identifying the exact patterns and signals that AI foundational models generate versus those that human generated content creates. And it varies for each different modality and we work on all of them.
- [05:03]
B
That's amazing. Thank you for sharing that. You know deep fakes video images you've mentioned soon to be text messaging, which is I think is pretty cool. All the top platforms obviously have their certain patterns and the way they create their generations, but your algorithm analyzing these images, frames and audio takes that to another level to really discern what's fake and what's not. As, you know, humans, it's easy to understand. Humans are so unique, but also they make a lot of, I guess, mistakes along the way. That's one of the things that helps us detect human activity. But I really appreciate you breaking that down for us, our audience today. Tolle, the next question I have for you is what industries do you see as most vulnerable to AI driven threats right now and how are they responding?
- [05:44]
A
Yeah, I'm happy to. And I'll cover this one from both a long term and a short term perspective. In the short term, I think it started out with news. I think that a more recent example is the protests in Turkey that were very real and very important to that country. And what was happening was there were AI generated images depicting scenes that actually never occurred. So it actually took a look away from the message. But news outlets were actually reporting on those AI generated images because they were much more, you know, social media friendly. They were much more click worthy because it had Batman and Pikachu and Joker and all these amazing dress up characters and protests which actually never happened. And I think for news and the information that we get, it's really important to discern what's real, what's not, what's misinformation and what's actually real. This was a political protest, but when you think about elections and other ways that people who gather information and make decisions, being able to tell it's actually AI or not is quite important to be able to do so. And when you have news outlets reporting on actually AI generated content, it becomes quite difficult for people to do so. I think after that it follows more like on the social media side of them being able to determine that and making sure that, you know, the information that's being shared is actually realistic and not a deep fake or not misinformation whatsoever. And there's actually countries like Spain and in China that are putting in laws going to effect this year where you have to be able to determine whether something's AI generated or not. And then finally, like more in the shorter term, I think our financial services industry is very much under attack. So whether it's AI generated scams or voice impersonations or KYC documents that are AI generated, all of those things are happening now with this new technology. And a lot of times you need to fortify existing systems to be able to protect against it. And we've seen sorts of crazy use cases too. We've even seen an insurance company who was getting AI generated X rays, dental X rays, to try to get insurance money out of, out of the company. So the use cases vary, but those are really the short term ones that keep coming up again and again and they're happening today long term. I think the Internet as we know it is very much endangered if you have an Internet where, you know the content that's being produced is AI generated, and some reports have it projected at 90% of all content will be AI generated over time. And all the comments that you receive are all AI generated. So what kind of environment is that when it's just essentially bots talking about. And we are very much keen to that and would like to play our part, which I think is a very important role, and to try to protect both information and, and against misinformation on the Internet as well as the overall world of what that looks like. Because I don't think any of us, you know, signed up to listen to podcasts of two AI speaking, which has happened before. Nor do we want to sign up and browse the Internet of all AI generated content and AI generated comments below it. So I think we're going to see a world where we might wonder, did we go too far with AI generated content? And AI or not, I think plays a really important role to try to protect against.
- [09:00]
B
Thank you. Appreciate that you covered quite a bit there. And yeah, I did see that about the protests in Turkey where AI was depicting images and frames of things that weren't really or didn't actually happen. You know, we saw this even non AI stuff where news outlets would pick up stuff. I remember during COVID were picking up stuff that weren't correct. Exactly right. And there was some, obviously some people suffered some of the consequences of that by not reporting or at least verifying the information. We're seeing some laws in China and Spain, as you mentioned, and I think that's important. But yeah, podcasts, gosh, if I could talk about that. Google's Notebook LM is pretty cool if you give it some prompts. It can actually do a podcast between two people, which is phenomenal. I think that's good and bad at the same time. So thank you. Tully, the last question of the day.
- [09:44]
A
I have for you.
- [09:44]
B
How do you view the future of digital identity in an age where AI can convincingly mimic real people, voices and behaviors?
- [09:52]
A
Yeah, it's a scary thought. And One that it's near and dear to me, having spent time with credit bureaus and KYC companies, you know, verifying millions upon millions of individuals. And I think a lot of the processes exist, are actually okay. Just need to be fortified against this new kind of dynamic of whether it's deepfakes or generative AI. I view this very much as like the new synthetic identity. Instead of, you know, spamming a credit bureau with data until a new identity is created, you're essentially providing different pictures or now videos of a really realistic person. And you're trying to convince someone, whether it's a platform or another individual on the line on the other side of the transaction that yes, this is indeed real. And there's a lot of repercussions to that. Whether it's money laundering, if a bank, let's say a fake person who's actually AI generated and now the money's being used for not so great things, or it's a, say a transaction on a marketplace, like even like Facebook marketplace, where the pictures on there are not real and the person behind it is not real. There could be really, really negative consequences there. And then even in the case we have this with the user who actually got defrauded for $150,000, where she was conned into buying fake art pieces and fake art because the artist created a bunch of AI generated pieces, sold them at a zone, even had a whole art exhibits with these pieces. And without checking one of our users who later found out that the use of AI or not was actually, you know, was actually AI generated. And I think this all starts with identity. If you can't find who the person behind it is, whether it's a financial transaction or a trust transaction like on the Internet, I think it has really, really negative consequences like financial and just trust on the Internet. And the example that you use like that, two AI speaking to each other, I think is a really, really cool example. But I also don't think that's what, you know, people want to listen to on their drive home is just AI speaking to AI when they sign into social media, I don't think they want. No one wants to see AI generated content with AI generated comments behind it. I would like to, I don't, for one. And I think there's a lot of negative repercussions of if we, you know, if we allow the world to become like 90% AI generated. So, and I think a lot of it starts especially with transactions and conversations. It start with digital identity. And I like to think I have been playing a positive role in it and continue to do so.
- [12:21]
B
Thank you. I appreciate it. Digital identity is going to be so important going forward. We're just seeing this proliferation of generative AI across every industry and platform. And again, I know there's a lot of great intentions, but they say the road to hell is paved with great intentions. Right. I just really appreciate you highlighting the stuff that you're doing today. I know we've got good processes, but we still need to continue to fortify and keep up with the advancement of generative AI. What I'd like to say, Toli, is I really appreciate your time on the show today, and I can't wait to talk to you again.
- [12:51]
A
Brian, thank you so much for having me. It was a pleasure.
- [12:54]
B
Bye for now.