AI: The AI Detection Tool Arms Race

0:00This message comes from Workday, the enterprise AI platform with a deep understanding of your organization's context and guardrails. So every AI action is permission-aware, giving you the ability to get work done right. It's a new workday.

0:22Been on the internet lately? Of course you have. Done some shopping, checked the news. Yeah, sure. Gotten fooled at least once by AI? More than likely. And when you got duped, you were not alone. I fell for AI this morning, yeah, yeah. And I literally accidentally reposted it. I don't even think I meant to repost it, but I did. And I'm getting DMs being like, girl, this is AI and you fell for it. It's over. I fell for my first AI video today. Little cute bunnies jumping on a trampoline. It's cooked, it's over. There was 13 million likes on it. 13 million people also fell for it. Then I showed it to my little brother and he's like, you know that's AI, right?

1:02No. No, I didn't know that was AI. And now I feel like I'm going to be one of those old people that get scammed. One poll by the data security company All About Cookies found that 77 % of Americans say they've been duped by AI content. Sometimes it can be harmless enough, like that viral TikTok video of the bunnies jumping on a trampoline. Now it's racked up over 25 million likes. But AI has also been used maliciously, of course, creating disinformation in politics, in medicine and health, in breaking news events, and across the culture. AI is also faking out people in creative fields from music to visual arts.

1:41And in education, teachers are raising the alarm about students using AI to cheat. Detection tools are out there to root out what's real from what's artificially created or enhanced. But can detection keep up with deception when AI is changing and improving so quickly? I'm Todd Zwillick, in for Jen White, and you're listening to the 1A podcast today. How can we tell AI apart from genuine content? Are we entering an era beyond verifiable reality? And what does that mean for our public trust and even our relationships? That's right after this. Stay with us.

2:23This message comes from Ali. Looking for more targeted support from your supplements? New Ali Precise Probiotics are formulated to go beyond general gut health, with options designed to support areas like skin, stress response, and metabolism. Get precise with your probiotics and choose support that aligns with your needs. With thoughtfully selected ingredients and specific benefits in mind, Ali helps make daily wellness feel more straightforward. Find Ali Precise Probiotics at a Walmart near you. This message comes from Betterment. Their automated investing and saving tools give you the quiet confidence of someone who knows where to put their money.

3:03With tax-smart tools that help grow your after-tax returns year-round. Get started today at Betterment.com. That's B-E-T-T-E-R-M-E-N-T dot com. investing involves risk performance not guaranteed betterment is not a tax advisor nor should any information herein be considered tax advice please consult a qualified tax professional let's get into the conversation with our guest with us is claire leibowitz the director of ai trust and society at the partnership on ai they're a non-profit coalition of academic civil society and industry leaders committed to the responsible use of ai she's led their work on Gen AI and deepfakes for almost a decade.

3:46She's also a PhD candidate at Oxford University and a member of the AI Advisory Board of the International Foundation for Electoral Systems. Claire, thanks for joining us. Thanks for having me. Matthew Stamm is here also. He's a professor in the Department of Electrical and Computer Engineering at Drexel University, where he leads the Multimedia and Information Security Lab. Professor Stamm, welcome to 1A. Thanks so much for having me. So very happy to have the two of you. Claire, help us understand the scope here. How often people are being duped by AI these days? Are they duped by certain kinds of content over others?

4:29So that's a great question. And whether or not slop has AI or it doesn't is kind of hard to gauge. AI is in everything. Even this audio recording is being compressed potentially with AI tools. So the question of whether or not there's AI involved or not is not really the right one. The question is, how much is there fake material on the web? And how much are these new, easy-to-use tools that help us simulate reality or even cast doubt on real images being deployed widely? So we've had the capacity to manipulate and create slop for a really, really long time. Whether or not that's being fueled into what degree is what's kind of harder to measure.

5:12So yes, we're seeing much more scale and the capacity to generate and this broader sense of distrust and confusion that will make people cast doubt on even real imagery. But it's really hard to quantify. But it's definitely way more out there. And the kind of monetization and speed and scale at which things go viral is making it even more complicated to catch up with. You know, Claire, that montage we played at the top of the show was a bunch of young people, and it was designed to show that, no, it's not just one of those old people that gets duped by AI. Those were all young people complaining that they got duped.

5:51And I think the message there is that this is pervasive at this point. Yeah, I think that's right. And it's not about age or expertise. The reality is these tools are so sophisticated that the things that maybe Matt and others would have said, like the hands looking strange or some of the artifacts, even at the technical level that media forensics experts like him use, we can't rely on those anymore because the technology has gotten so good at simulating reality. So even I, who've worked in this field for many years, when that pope in the big marshmallow-like puffy jacket a few years back was going viral, that was plausible enough that he, in Italy, a very chic nation, was going to go around wearing that big puffy jacket.

6:34And even experts, the naked eye, whether you're young, whether you're expert, these tools are so good and they're embedded in all of the media almost as a recipe that we're seeing today. So Matthew Stamm, talk about that. And Claire talks about, you know, it wasn't long ago, yes, that what you would look for is AI didn't know how to make hands. So you'd see a photo, it would have either eight fingers or the fingers were weird, or they just didn't look quite right because you're human and you know what a hand is supposed to look like. It's not the case anymore. I mean, AI has gotten much better at hands.

7:07It's gotten much better at a whole lot more. What's your sense at the speed at which AI is becoming just harder to detect now? If you're talking about detecting it by looking for some visual inconsistency, I no longer trust my eyes. Generative AI has gotten so good. We can't rely on seeing things like that extra finger or a weird texture in the background. Or even with video, in the past, maybe about a year ago, you'd see odd motion physics, but that's starting to look very realistic. Instead, the way that researchers like myself work on detecting this is not looking for visual cues. It's looking for statistical traces that are left inside the image or video or audio signal by how the media is formed, either in a camera or in a generative AI system.

7:59So we've seen so much money and research put into this technology that it's advancing at a pace that's pretty unparalleled. And this can be a really great thing because AI can enable a lot of really powerful tools for good. But these tools can be misused to create misleading media. The types of detection tools that you're building, Matt, again, they don't try to detect eight fingers. You can't do that anymore. They're really looking in the code. They're looking behind the scenes of the image. We're talking about maybe an image in this case, but there's other forms of AI too. Is there a form of AI that they're better at detecting versus others?

8:40Are they better at spotting a fake video than an AI-generated song? Does it matter if you're a detection computer? Yes. So the way we think about it is different ways that you can create or falsify media require different tools to detect them. So even within something like an image or a video, there's no one single approach to detect everything. At a high level, though, the way we think about it is when you create a piece of media, let's stick with images, whether it's in a camera or in a generative AI system. The processing inside the camera or the generative AI system leaves behind these statistical traces similar to how a criminal leaves behind fingerprints at a crime scene.

9:25We create a set of tools that look for these fingerprints, figure out how to pull them out and analyze them. The tricky thing is fingerprints left by an AI image generator are very different than fingerprints left by an AI video generator. Or fingerprints left by Photoshop look very differently than fingerprints left by something like ChatGPT's image generator. So we have to make a lot of tools to test for a lot of different ways media can be fake. Yeah. Claire Lieblis, what are some of the other issues that these detection tools run up against? I mean, me, a layperson is thinking of one right away, which is if the AI generation leaves behind fingerprints that the camera doesn't, you can write AI code to fix the fingerprints.

10:15Is that true? So I think when we talk about first detection, not to get really wonky, but being able to gauge if something has been augmented, there are a lot of different technical methods for doing that. One is looking, as Matt described, at all these different artifacts or things in it. And one is actually called fingerprinting. So Matt used fingerprint almost as a way of colloquially describing it. But there are other ways to almost insert earlier on with an actual kind of pro-social goal, not just hoping that you can glean something from maybe a bad actor, actually injecting that fingerprint from the point of creation of the content so that you can have it travel over time.

10:57So different techniques have different susceptibility to manipulation, removal. So it depends on the technique. So when we say detection, that's kind of almost assuming that nobody added anything. You're just looking at the typical laws of physics in an image that you want to gauge from. What I will note, though, is detecting whether or not something has been augmented with AI says nothing about whether it's actually malicious or deceptive. And the example, and that's actually what matters to people. So there's one question, which is how good are these technical methods for even assessing the presence of AI or not?

11:35And then there's the question of how do you convey that degree of probability to a layperson who's working at the speed of the web and ultimately needs to decide, do I trust this or is it true or is it false? I just wanted to ask Matt, Stan, before we go to a break, you developed some of these tools in your lab. It's obviously an industry. We're in an arms race. I'm just going to, that's what it sounds like to me, detection versus deception. What does the detection industry look like right now, Matt Stam? How fast is it growing? It's emerging very quickly. We are seeing a number of companies just starting.

12:13Right now, a lot of the leading work is coming out of academic labs, and there's some early companies working to transition that out. We'll see more in the future. Stephen emails this. The answer is actually very simple, but AI companies will never willingly agree to it. We have to require all AI output to come with a digital signature that identifies it. Then we don't need fancy detection tools at all. Stephen, good point. Would AI companies ever agree to it? Would an independent actor using AI agree to use a watermark? I doubt it, but it's a great suggestion. We're going to talk more about how companies are using AI detection in social media and in universities and other places as well.

12:56It's coming up right after this.

13:01This message comes from Ali. Looking for more targeted support from your supplements? New Ali Precise Probiotics are formulated to go beyond general gut health, with options designed to support areas like skin, stress response, and metabolism. Get precise with your probiotics and choose support that aligns with your needs. With thoughtfully selected ingredients and specific benefits in mind, OLLI helps make daily wellness feel more straightforward. Find OLLI Precise Probiotics at a Walmart near you. Support for this podcast and the following message come from Strawberry.me. Be honest. Are you happy with your job?

13:40Are you stuck in a job you've outgrown or never wanted in the first place? Are your reasons for staying really just excuses for not leaving? Let a career coach from strawberry.me help you get unstuck. Discover the benefits of having a dedicated career coach in your corner and get 50 % off your first coaching session at strawberry.me slash NPR. We've just marked one year since Congress voted to eliminate federal funding for public media. NPR remains committed to informing the public, but a free press doesn't just happen. It's something we need to protect. Without federal funding, we are relying on your support now.

14:19Please show your support before the end of the month. Visit plus.npr.org. Let's get back to our conversation now with this message that we got from one of you. I feel that any time we hear or read AI-generated information, it should be identified as such. should have a little disclaimer somewhere that says AI assisted or AI generated. And I think that that should be across the board. Yes, a couple of you have had that suggestion. Stephen had the same suggestion before the break. Let's bring in Hadass Gold, AI correspondent at CNN. Hi, Hadass. Hey, great to be with you. Well, you heard the suggestion from our listeners just there, Hadass Social Media Platforms.

15:09They've been toying with labeling AI content. Meta had an announcement in 2024 that it would start labeling content that it indicates was AI generated. It's been, I think, criticized for inconsistency, false labels. It's not perfect. What's the state of AI labeling, at least on social media? How good is it? How helpful is it? So pretty much all social media platforms these days have some sort of AI labeling where it applies a visible tag that might say AI info, which might be anything from AI was helped use to edit it to something was completely AI generated. And all of these different platforms use some variation of either they ask the user to apply a tag when they've used AI themselves, but they also use hidden industry standard AI markers within the files themselves, meaning that they have systems that will scan what you upload to try and figure out these sort of within the data that can signal to them that this was inauthentic in some way.

16:18but they are not perfect as we've seen. And there are also, you know, users are getting good at finding ways around this. And I myself have seen plenty of content that I am nearly 1000 % sure is AI generated that does not have a label on it. And you just have to be aware that these labels are not perfect and that obviously labels are helpful. Sometimes it doesn't even matter. People might kind of not see the label, not care about it, but just know that, you know, Even if something doesn't have a label on social media, it doesn't mean that it is for sure 100 % authentic content. Yeah, I mean, Hadass, just in the last week, you saw this one, I'm sure.

16:57I saw a video that depicted President Trump trying to get into his limo, the beast, and then like collapsing into a heap and having to be helped up by a Secret Service agent. Looked totally real. It was completely fake. I mean, I saw it 50 times on various feeds. It didn't have a label on it. Yeah, and that's what's really difficult about all of this is that sometimes being something being AI generated doesn't necessarily mean that it is from scratch AI generated because you can use AI editing for actual videos. Yeah, and I think in this situation they did because there was a companion video of, yes, Trump getting into the beast that was just like it, but it was modified.

17:38Exactly. So they just edit and modified it. And, you know, there's a lot of technical aspects, which I'm sure your other guests can go into, of maybe why the tools, the scanning tools might not catch those or how that all works. But it's a cat and mouse game for these platforms to try and catch up and try to label everything. You just have to think about the sheer amount that is being uploaded at any one second on all these platforms. And, you know, people just need to be kind of smart consumers and not believe everything that they see. Just like if you were watching late-night commercials and they would be like, as seen on TV, look at this amazing product and what it can do.

18:13I think people realize that not all of those products do exactly what they can do, and not everything that you see on social media should be taken as real. Not everybody buys those products, can do everything they say. But again, those products make millions of dollars before the infomercial is over, so it's a conundrum. Let's go back to our voicemail. Here's Jesse in Kansas. There are always going to be tips coming out about how to spot AI, but it really doesn't matter because it's always just going to be improving and changing. So Matthew Stam, what about that? I've already sort of used the term arms race.

18:47I don't know if it's quite appropriate, this battle between deception and detection. But Hadass just described this Trump video, you know, a combination of reality and AI that didn't have a label. And Jesse says, look, this is moving so fast, we're never going to quite get there. What does that dynamic look like? Yeah, so I hesitate to say it's an arms race because an arms race means that the AI companies are trying to actively outgun detection companies. In several cases, they're actively working with governments and other folks to try to make what comes out of their generators more detectable.

19:30The real challenge is that generative AI is one of the hottest research topics in the world. It's viewed as a major economic driver, and the amount of money and talent working in this area is making progress in that area so fast. It's nearly impossible to keep up. At the same time, we're still thinking about this in kind of an old way of information security. So I think that labels and watermarks are really helpful tools, but this is all assuming that the information that you're getting comes from a known and trusted party that's going to put in a watermark or a label or some external security measure.

20:09But I don't think that's the way that we encounter information nowadays. Much of the time we see information from an unknown or untrusted party. It might be AI generated, but it might not be AI generated by a Western tech company like Google or OpenAI. They don't have any requirement to buy into trust or provenance ecosystems, but we still have to contend with that media. So we're seeing the emergence of a really new information security problem, and we're still scrambling to catch up with this, not only technically, but also socially. Let's talk about the social problem, Claire. I mean, Matt puts his finger on it.

20:48Everybody stipulated that a lot of times you can no longer tell what's real. Claire, you're in the business, and you said you – Matt, it was you who said you don't even trust your own eyes. So where does this go, do you think, Claire, socially? Education is important, what to watch out for. Remaining skeptical of everything that crosses our screen has a big social cost of public trust, doesn't it? That's absolutely right. Right. And this is a term we've been thinking about for many years called the liar's dividend, which is the idea that just the sheer possibility of augmenting content with AI and the ease of that makes it super simple to just use plausible deniability to cast doubt on video evidence of human rights abuses or actual political crowd sizes, as we've seen in the past, or just to say that real content is fake.

21:39So this broader social distrust in real media and beyond is something we have to be really mindful of. And all of the examples that we just heard are going to actually be mandated. The European Union's AI Act is going to affect this summer, and it mandates AI labels across all of the platform infrastructure that Hadassah was describing that they've already adopted. But what we've seen in some of our case study work with the companies at the Partnership on AI is actually that if we label everything, it's just going to be so ubiquitous that it's going to mean nothing to people. Google reported that the more they labeled content as best they could and adhered to these kind of norms, actually people were more trusting of unlabeled content to this point that you can't just assume that's da-da-da-da said that just because something's unlabeled mean it's automatically truthful.

22:27So I worry about kind of as the technology is just so enmeshed in what we're seeing, the skepticism being not just healthy skepticism, but over skepticism. And also some companies are resorting to just placing salient labels on high stakes contexts like health or civics, because they don't want people to ignore these labels. They want them to be present, but worry that if they're on everything, we ultimately will just create a degree of distrust that actually backfires and makes people confused. Hadass Gold, I'm interested in how you're thinking through this because of where you're situated in the world.

23:03You cover AI, of course. You've also covered politics in the past around the world and here in the United States. You work for a major television network. You're in on editorial meetings with your bosses. How are you thinking through and what are editors saying about verifying content, knowing what to believe, how to report on it? You've got a problem sort of going both ways, it seems to me. Yeah, we would never report on something that we saw on social media without going directly to the sources and talking to people who were either physically there or who took the video. We have a whole team here that works to verify these types of videos, especially when we're talking about disaster zones or conflict zones.

23:52We do not take anything that's been on social media without going through a meticulous process of geolocating everything, and even then just being very careful of what we're seeing just in case anything has been edited. And the best way, you know, as journalists to get around this is to just use reporting and pick up the phone, call people directly, you know, actually talk to them and make sure that you're getting the story from multiple sources, multiple angles as much as possible before you report. I think we would never report on a video, for example, of President Trump, you know, falling with and, you know, thankfully he has a press pool with them without talking directly to people there.

24:31You know, this is a big deal during the assassination attempt on President Trump, where there's a lot of theories and stories flying around. I mean, there's videos going around, and you have to stick with, you know, the actual, you know, eyes on the ground who can tell you what they saw. Well, we got this message from Ingo, who says, what if I created an AI image, let's say, of a senator in the hospital? Then I print a high-quality print of this fake. Now I take a picture of that image, can we detect that? Matthew Stamm, you're smiling, and I think you're smiling because Ingo is making a reference, I think, to Senator Mitch McConnell, a recent, I'll call it a proof of life photo that was published because he's been in the hospital, he's been sick, and he hasn't been able to return to the Senate.

25:18And you have been involved in efforts to detect with the news media, whether that photo was indeed real or fake, because the amount of, I can't even call it skepticism, just the churn of conspiracy theorizing around that photo was instant and was right across social media when it came out. Yeah. So I was asked to help verify that image by a couple different news agencies last week when that came out. What we found from looking at that is we did not find evidence that that was AI-generated or AI-modified. A couple caveats, though. When people like me or commercial entities examine something, the best we can say is we haven't found evidence that something's fake.

26:09It doesn't mean that it's real. It also doesn't mean that we can't figure out when an image is captured. For that particular image, it appears to be real. One of my colleagues also was able to trace down the newspaper that Mitch McConnell was holding. And it was, I think, a recent copy, maybe that day's copy of The Washington Post. So we're pretty confident that that's real. One of the big problems you brought up, though, is even when something is real, there's so much speculation around whether something is AI-generated or not. sometimes it doesn't matter if something is real. You've already undermined trust in the information that you see.

26:51Really, what we're seeing is a big change in the way we view information. And that's come about, I think, because of how quickly we've transitioned from an image being real capture of reality to something that's really malleable by just about anyone with no particular technical skill. Yeah. Hannah writes in to say, I'm concerned not only are people being fooled by AI, people are also able to use that distrust of what we see to dismiss what's real. Claire, that's what you called the liar's dividend. Claire, what did you see in the McConnell proof of life photo story trying to verify? Matt makes the point because we didn't find evidence that it's fake doesn't necessarily mean it's real.