To Spot AI Fakes, Think Like a Digital Forensics Expert

[00:00:00] I want to start with a post. One that, just a few weeks ago, popped up on my feed over and over again. And probably yours too. It was around the start of the World Cup, weeks before the final match, in which Spain beat Argentina in extra time.

[00:00:22] A nearly decade-old picture of the competition's most prominent players circulated online, showing Argentine legend Lionel Messi. At age 19, bathing a baby. That baby's name? Lamine Yamal, the Spanish superstar prodigy. No one could have ever imagined that the child in the photo shoot would one day end up opposite Messi in a World Cup final.

[00:00:52] Seriously, no one. It was so unbelievable that the photo sparked the same reaction in a lot of people, including New York City Mayor Zoran Mamdani. As someone who likes to think of myself as technologically fluent and able to spot AI versus a real image, I thought that the image of Lionel Messi and Lamine Yamal was AI. I don't know if you've seen this image.

[00:01:17] But it was real. In 2007, Yamal's family won a raffle. And the prize was a photo shoot with then-FC Barcelona player Leo Messi. As we've talked about the past two Saturdays, even real facts are a casualty of the rise of AI. There's so much of it in our social media feeds, it's causing us to question everything. And that's having an impact on how we see the world. I'm Nicole Nguyen, personal tech columnist at The Wall Street Journal.

[00:01:47] It's Saturday, August 15th. And this is the third and final episode of our special series, AI and the Blurring of Reality. Today, we ask the experts how they figure out if something is AI. And show you how to think like a researcher, too. Major news moments are a big opportunity for people to use generative AI

[00:02:11] to quickly pump out full-blown scams, memes, and engagement bait. One TNP listener, for example, told us about an online betting company in Argentina that used the likeness of the late soccer player Diego Maradona to create an AI advertisement. Maradona passed away in 2020.

[00:02:40] And our listener says he worries that not everyone is aware of that. Manipulating media may not be new, but the sheer volume of content and the pace of distribution is. We've obliterated barriers to entry for anybody to enter into this space. That's Hani Farid, the digital forensics expert from the first two episodes. If you haven't already, give them a listen.

[00:03:09] Trying to spot AI fakes isn't easy. And it's getting harder every day. It's nearly impossible for most people to verify if a post in their feed is real or not. They can't. There's no other answer to that question. And they can't for a couple of reasons. One is, this is hard. I do this for a living and I'm pretty good at it and it's hard. Number two is, even if I told you certain things you can look for, this is too fast of a moving space for me to say three things to protect yourself online.

[00:03:39] Don't read that article. It is utter and complete nonsense. In fact, it's worse because now you feel like you have a false sense of security. For Farid, there's a very simple solution to all of this. Stop getting your news and information from social media. If you want to go there and be entertained, I don't care. If you want to go there and connect with your friends and family, I don't care. But don't, for the love of God, use these mediums as a place to get news and information about what's happening in Iran or Gaza or anywhere.

[00:04:09] It's not what it was designed for and it's not what it's good for. But there's good news here because we work with news organizations. We help fact checkers. You want reliable information? Go to the people who do this for a living. We have to go back to trusted sources. That may be easier said than done. Billions of people, nearly all internet users, use social media, according to the latest data from consumer research firm Kepios.

[00:04:35] So we asked Farid about how a specialist goes about auditing posts. And what are these forensics experts doing? Are they, you know, zooming in to the micropixel level to see whether or not it's synthetic? So there's some tension here in your question. Because what you're asking me to tell you is how do I defeat my adversary while my adversary is listening to this conversation. And that doesn't seem very advisable on my part, does it? But I'll give you some examples.

[00:05:06] What you have to understand about generative AI is it doesn't understand the physics and the geometry of the physical world. It's a statistical inference engine. It has taken in, digested billions of pieces of content. And it has learned what a natural image looks like. It has learned what a video looks like. It has learned what a voice looks like. But it doesn't know anything about cameras or optics or sun or geometry or shapes or structures or reflections.

[00:05:32] Basically, AI looks for patterns, which it's good at. But sometimes those patterns don't match up with the real world. And it has a harder time with that. And so what that means is it often gets the physics of the world incorrect. So, for example, if you go outside on a sunny day and you take a photo with lots of peoples and cars and buildings, there is one dominant light source, the sun. And it will cast shadows. But it casts shadows in a very specific way that have to be physically consistent.

[00:06:02] And the AI just gets that wrong. Is the geometry, I'm looking at your background. You are in a room where the wall is receding away from you. There is perspective distortion. Things get smaller as they go further away from you. But in a very specific geometric and mathematical way, AI gets those things wrong so we can measure those things. Beyond those, Freed says there are other tells that digital forensics experts such as himself have honed in on.

[00:06:29] Other patterns that AI does or doesn't add that can reveal a piece of content was AI generated. But understandably, he didn't elaborate more on those with us. And is it only a matter of time until these models figure out things like perspective and dominant light? You know, like as an art major, I learned that the Renaissance painters learned about perspective and could finally draw feet with the correct depth and had that revelation

[00:06:57] and that allowed the Sistine Chapel to look like as it does. So will these tools and your methodology have to evolve also? So two-part answer to that. One is almost every tool that we have developed has a limited shelf life. It will work for a while and it will stop working. And then you need a new tool for the next generation. That's the way cybersecurity works, right? We adapt, the adversary adapts. We adapt, the adversary adapts. But the other part of that answer is really interesting to me,

[00:07:24] which is OpenAI and Google and MidJourney, they're not my adversary. They're not trying to defeat me. Sam Altman in OpenAI, he's not trying to build an image model that can defeat forensic analyses. What is he trying to do? He's trying to build an image model and a video model that is appealing to his customers. This is not like a classic cybercriminal defense relationship where the cybercriminal really is trying to defeat me. But it's one thing for an expert to spot a fake by analyzing pixels.

[00:07:54] And another for the rest of us to come across that fake on our phone during our morning commute. After the break, we'll look at who is most vulnerable to these scams. And why even veteran fact-checkers admit that being fooled is part of the job. When thinking about which demographic will be most impacted by fake content,

[00:08:24] research shows some groups will be hit harder than others. Here's digital forensics expert Hani Farid again. Older generations are having more trouble. My parents' generation, soon to be my generation. They're not used to it. They're used to, oh, I read it, therefore it's true. I mean, honestly, say what you will about Gen Zers and millennials, they're pretty savvy. They grew up with this. That may increasingly not be the case with the youngest folks, by the way.

[00:08:51] A 2025 report by the nonprofit group Common Sense Media found that teens are now regularly misled by fake content online. We have looked specifically at the belief in, this was a few years back during the times, of COVID mis and disinformation. And we looked at age and political affiliation and gender and geography, and there was not a lot that correlated. There was these sort of nuanced things, but it was across the boards. You saw people believing things that are untrue.

[00:09:20] Left, right, middle, center. Everybody's vulnerable. Even some of the most seasoned fact-checkers. I have to admit, I was initially fooled by that Pope photo in the jacket. That's Megan Peterson. She's the deputy photo director at the Journal, and a big part of her job is to verify the legitimacy of photos and videos shared online. Take the 2023 viral post Megan was just talking about. You might remember the image, which showed Pope Francis

[00:09:49] wearing a swaggy, oversized, white puffer coat. It looked really real. And, you know, I saw it in passing on your smartphone, which is much smaller, so I'm not looking at it in detail. So I did get fooled by that one at first, yes. Me too. I wanted to believe that it was true, I think. Yes, also that. And that might be part of the problem. Some of these images, they provide some lightheartedness or some enjoyment, and you do want to believe that it's something real.

[00:10:17] But it does make it a little bit more difficult, because you're not necessarily using that critical side of your brain when you're looking at that, which is okay. We don't always have to have our critical side of our brains turned on. But in some ways, it might make it more likely that you are going to miss something important. I think it's in some ways desensitizing people to have all this fake content out there. When it comes to solutions in this new frontier, there aren't many yet.

[00:10:46] One way to tackle the issue is AI labels. Platforms like Meta and TikTok have put in place guidelines and created tools to manage AI content. But critics have noted that the labels are voluntary, and the tools aren't perfect. A number of creators, for example, have complained that their non-AI work was mistakenly labeled as AI. Some companies have also begun applying visual watermarks or digital metadata to signal that a piece of content is AI,

[00:11:16] but those markers can be removed or edited. We reached out to several platforms to ask how they're tackling the spread of AI posts and deepfics. Meta told us the company is continuously improving its labeling systems. YouTube said it mandates AI labels, automatically tagging AI content, and offers a likeness detection tool to search for deepfics on its site on behalf of creators. TikTok did not have a comment.

[00:11:43] Okay, let's get back to what you can do. On our social media feeds, we're faced with a spectrum of manipulation. Some posts look way more AI than others. For ones where you're not so sure, here's how Megan inspects questionable posts. The easiest way is really to zoom in and really look at the image. You will often find some anomalies there. One good example is they're not very good at rendering hands or limbs.

[00:12:12] There's an image that came out last year from the Trump-Zelensky meeting. There's an image of European leaders sitting in the hallway outside of the office that was shared around social media. And it really did, at first glance, look like a smartphone photo and something that was real and happening in real time. But when you zoom in, you'll notice that there are more legs than people. So you can start to see these things when you really just zoom in and look at an image. They're not always obvious at first glance.

[00:12:39] You'll also start to see there's certain things we know about objects. Like an iron beam is straight. So if there's a crane in an image and the lines of the crane are wavy, it's likely not to be real. You can also Google those objects and see what they look like in real life to start to get some comparisons and some ideas. Something else we'll do is verify from other images of the scene or Google Street View. Try to compare that with real life and other known images.

[00:13:07] And beyond zooming in, Megan also checks the context of the scene. Looking at the weather on a specific day is also really helpful. If it's supposed to be pouring with rain and the image is sunny, that's an indication that it was probably either not taken on that day and is being portrayed in false light or was made up entirely with AI. So you really just have to zoom in and look and just look for things that don't feel right to you.

[00:13:34] I also wanted to know, what about AI detector tools that popped up shortly after ChatGBT did? Such as Zemble.ai or AI or not. Google also incorporated a detector into Gemini. You can upload an image and ask Gemini, was this created with Google AI or is this AI generated? I asked Megan about those types of tools.

[00:13:59] There's a lot of AI detectors, but there's no one single AI detector that really gets it all right. We've tested a combination of them and have found between two of them, you can catch a lot of things. But there's false positives and false negatives. That is really my biggest tip is to trust your gut. If you're not trained in photography and lighting, you might not know it's the lighting, but you will feel that something's off.

[00:14:23] Trusting your gut may work for now, but as these AI tools get even more sophisticated, how long can that really last? I asked Hani Farid. I know that academics don't like to speculate. But if one were to speculate, where do you think this AI manipulated image future is headed? Well, there's a few things I'm comfortable on speculating. Yeah. So one is the images will get better, the audios will get better, the videos will get better.

[00:14:53] And by better, I mean more visually and visually compelling, right? Higher resolution, higher quality, longer videos. That trend will continue. More and more people will use these technologies for good, for creative ways, and for nefarious ways. I think that's almost certainly the case. The question of are we going to be defeated, if you will, is an interesting one. And I think that's what you're really asking. So I think the answer is yes, but with a footnote.

[00:15:19] If you think about when I leave my house in the morning to go to work, I lock the door. Does that prevent a skilled burglar or a locksmith from breaking in? No. But does it prevent 99.99% of the people out there? Yeah. And I'm comfortable with that, right? My job is not to prevent somebody from creating a deep fake. My job is to prevent 99.9% of people from doing it. You mitigate risks, not eliminate them.

[00:15:47] And if I can mitigate this risk, that is that the average person on the internet trying to poison the information ecosystem on social media cannot get away with it, but very sophisticated, well-funded, well-resourced actors can, that's a risk I can live with. And I think that's probably we will end up in some steady state like that in some single-digit number of years. If you're looking for a silver lining, maybe it's this. AI will get better, but this Wild West phase won't last forever.

[00:16:17] Most technologies follow a similar pattern. Not a lot changes, not a lot changes, and then it's a hockey stick. And then there's radical change, and then things start settling down. I mean, if you think about these mobile devices in our hands, you know, there was a time when there was a lot of advances happening, and now we're more or less steady state. We know what to expect iteration after iteration. So, are we doomed? Maybe not. Or maybe not yet.

[00:16:44] But we are in the steep part of the curve Farid is talking about, where the technology is advancing and adapting so quickly, our brains are having a hard time keeping up. Having spent the first half of my career covering the last great technological leap, the smartphone era and the appification of everything, there's one thing I'm optimistic about. When the tech becomes mainstream, we all get savvier about not clicking that suspicious link.

[00:17:10] Over the next few years, we'll have to steel ourselves with that same vigilance, to rely more on trusted sources and not believe everything we see on the internet. Though, too much skepticism might erode our shared sense of reality. To avoid societal confusion, despair, cynicism, we'll need to figure out how to strike the right balance, and who to trust for information. If those past tech cycles are any indication,

[00:17:39] that could take some time. And that's it for the third and final installment of Tech News Briefing's special series, AI and the Blurring of Reality. Today's show was produced by Julie Chang. I'm your host, Nicole Nguyen. Michael LaValle wrote our theme music and mixed this episode. Our supervising producer is Katie Ferguson. Our development producer is Aisha Al-Muslim. We had additional support from Wilson Rothman.

[00:18:08] And Chris Inslee is the deputy editor of audio for The Wall Street Journal. We'll be back Monday morning with your TNB Tech Minute. Thanks for listening.