A test to weed out AI-generated deepfake images
06:58 Share

This podcast is brought to you by LHH, the Global Talent Solutions and Advisory Company. What does work really mean? For many, it's just transactional, functional. But LHH believes it can be more. Work isn't just about tasks and deadlines. It's about passion, people, and possibilities. With the right guidance and vision, incredible things can happen at work. Finding the perfect hire, nurturing talent, making the ordinary extraordinary.

LHH doesn't just find beautiful moments at work. LHH creates them. Recruitment, development, career transition. LHH, a beautiful working world. Learn more at LHH.com slash beautiful. How good are you at spotting AI-generated fake images? From American Public Media, this is Marketplace Tech. I'm Novosafo.

AI-generated deepfakes are everywhere on social media, and you can take a test developed by Northwestern University to see how well you spot them. I took it myself, sifting through a bunch of real and fake images. Okay, now it's two guys sitting at a park bench looking at their phones. I'm going to go with...

I was right. This is a synthetic image. So I got five out of six right, which is the average in a study Northwestern conducted. Lead researcher Matt Groh also helped develop a litmus test, a series of things to look for to spot deepfakes.

The taxonomy here is anatomical implausibilities, like way too long neck, you know, six fingers kind of thing. Stylistic artifacts. This is when it's too waxy or too shiny. Or functional implausibilities. That's like where the guitar strings or the tennis racket strings are not taught like you'd expect them to be. Or any other kind of thing around that. Violations of physics. Or

where the shadow doesn't match up with the other shadows in the scene and what you expect the source of light to be. And then sociocultural implausibilities, where it just doesn't make sense with whatever the social or cultural or historic context might be. And you found that some of these taxonomies, some of this criteria, if it's in the photo, is easier to

to kind of spot and say, oh, this is a fake image versus some of the other, which is the easiest one to spot, which is the hardest one to spot? This, of course, depends on the person. But, you know, the anatomical implausibilities are the clearest in many ways. And it's because, you know, we're used to seeing lots of people. And so when you see an extra leg appear in the image, you know immediately that that image is unrealistic. At

At the same time, it might not always be so obvious that there is three legs in the image. And so I can show you a basketball player image where if you look for 15 or 20 seconds, you can see the weird leg that is appearing as a third leg, but it's actually not something that's immediately obvious. And so maybe the best way to think about this is...

If you played Where's Waldo when you were a kid, you experienced this thing of you looked at the page, you know Waldo was in front of you, but you couldn't find Waldo until you spent a few minutes and you had to kind of put a spotlight on Waldo and then be like, ah, that's Waldo. Coming up, we go from Waldo to Dumbo. Yes, the Disney character. We'll be right back.

This Marketplace podcast is supported by Greenlight. As a listener of Marketplace, you're likely already building smart money habits for you and your family, trying to instill important lessons on saving and spending and the economy overall, and the younger folks in your life. But what about the older generation? Your parents, grandparents, aunts and uncles? As they age, they may need more support in managing their finances too.

With Greenlight's Family Shield, you can take the next steps by protecting your senior loved ones from scams and financial fraud. Family Shield offers account monitoring, real-time alerts, and coverage, including up to $100,000 for deceptive transfer fraud and $1 million for identity theft, so you can keep your whole family financially safe and sound. And with added safety tools like SOS and crash alerts, along with location sharing, you can keep an eye on both your family's wallet and their well-being.

Take care of your whole family, from kids to grandparents, with Family Shield from Greenlight. Sign up today at greenlight.com slash marketplace. If you're struggling to keep up with all the latest innovations in tech and what they'll mean for your life, TED Tech has you covered. Get ahead of the curve with digestible downloads on some of the biggest ideas in technology, from AI and virtual reality to clean tech. Find TED Tech wherever you get your podcasts.

You're listening to Marketplace Tech. I'm Novosafo. We're back with Matt Groh, professor at Northwestern University, who studied how adept we are at detecting AI-generated fake images.

How much of this research and these kind of this, everything you just described, how much of that also applies to video that's out there? Because we're starting to see some very sophisticated AI generated video. Yeah, I know. And I think what you're referring to is VO3 that came out of Google. I think it all applies to video.

But it depends. Okay, so let me kind of like, think about this framework. So we talked about anatomical implausibilities. And in a image, you know, the anatomical implausibility has to be in that particular frame. In a video, the anatomical implausibility, if you just find a single frame, you know, there's about 22 frames per second. And so that means that over 10 seconds, you already have 220 frames. If there's a

single anatomical implausibility, that's going to quickly be able to let the human know there's something wrong here. And an example of this appearing is at the very end of our paper, we show a Sora video. You know, Sora is a model that OpenAI has. And when it launched, they showed a woman walking down the streets of what appears to be Tokyo.

And at second 16, what happens is her left leg becomes her right leg. In other words, rather than stepping front, back, front, back, or one step, two step, they kind of rotate in a way that I've never seen before. And I know that no one walks like that. And it's just...

physically impossible. So maybe that's a combination of, you know, a anatomical implausibility and a physical implausibility. But the idea here is that we're building a language for which we can quickly identify why something is likely to be fake.

Right now, if you're an average user trying to navigate all this and figure out what's real out there, what advice would you give people? The first piece of advice is to be a little bit more skeptical today than you might have been 10 years ago.

and to not just automatically trust everything and to sometimes double check things. You know, I'll give you a quote from Carl Sagan. Carl Sagan said, extraordinary claims require extraordinary evidence. And so, for example, I have an office right on Lake Michigan here in Northwestern. And if I told you I saw an elephant just fly by, you would say,

no, no, like I saw Dumbo, but that's not real. Now, if I just said, okay, well, let me show you the video of Dumbo flying by. And I shared the video that looked realistic, not cartoonish at all. You'd also say, Matt, like we're literally talking about VO3. Like I know that this can happen. Now, if I...

flew you out here and you saw it with your own eyes and experienced it, you also might like say that's not real. And then if I like had you parachute down onto the thing and you like gave it a like a little like ear rub and stuff, you're like, okay, actually, like I'm pretty surprised, but maybe this is true. So you really got to not trust your lying eyes. That's exactly right. And so I should say the not trust your lying eyes

You know, there's an adage, seeing is believing. But that is the first part of the adage. There's a second half. The second half is, but feeling is the truth.

Seeing is believing, but feeling is the truth. And here, feeling is not just your emotions in general. It's experiencing the situation in its fullest content. So every single sense that you have, also connecting that with your past experience and your memories, that is all part of whether you identify, like how you identify the truth. But seeing is just one of our many senses for getting there.

That's Matt Groh with the Human AI Collaboration Lab at Northwestern University. We'll have a link to Northwestern's study on our website, marketplacetech.org, along with a link to that online image test. Jesus Alvarado and Daniel Shin produced this episode. I'm Nova Sapo, and that's Marketplace Tech. This is APM. Hi, I'm Katie Drummond. I'm Wired's Global Editorial Director, and I'm excited to be joining the hosts of our flagship podcast, Uncanny Valley.

It's a show about the people, power, and influence of Silicon Valley. It's hosted by some amazing Wired writers and editors, where each week they discuss the influence of technology and culture from the Valley on our everyday lives. But we're also adding another episode to that feed, hosted by me. Each week, I'll have an urgent conversation with one of our extremely busy Wired reporters or editors about this week in news.

Our journalists are constantly asking smart questions to find out where they lead and to help you understand where the world is going a little bit better. I hope this new weekly episode does just that. Make sure you're following Uncanny Valley in your podcast app of choice so you don't miss an episode.