The Future of Deepfakes and the Decline of Reality with Hany Farid

0:01Hany Farid:This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome, that's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up required, compatibility and availability varies 18+. Forget about the bad guy creating the app. There's nothing you can do about that person. They're like the street-level dealers.

0:31You got to go after the kingpins. That's who you go after. And you bash them over the head until they get their act together.

0:59subscribers also get access to additional episodes where we respond to their best comments and they get early access to our interview series to gain access to that content at 4formedia.co also remember to subscribe to our youtube channel where you can watch all episodes i think we're pretty good at telling the difference between an ai generated image and a real photograph but when i need help i call hani farid which we have on the podcast today Fareed is the co-founder of GetReal, a company that specializes in detecting deep fakes and is the developer of PhotoDNA. You may not have heard of PhotoDNA, but PhotoDNA is a very important piece of software that today is used by any serious internet platform that automatically detects some of the worst images that exist online.

1:55And we started talking to Hani pretty regularly in 2017 when Sam first reported on deepfakes. Has that technology evolved and changed how we perceive reality? So have our conversations. I wanted to talk to him today on the podcast so you could hear one of those conversations. And so that Hani and I can zoom out, reflect on the past few years and speculate about where we might be headed next.

2:29Okay, Hani, thank you so much for coming on the podcast. It's good to be with you. Just to give some background to how we know each other, I think we started talking when we were still at Motherboard and Sam was doing some of the first reporting on deepfakes. Not some of the first, the first reporting on deepfakes. I was just saying the first. When we were trying to wrap our head around the problem and how tech platforms are going to deal with it, we immediately ended up looking at photo DNA, which is something you worked on. Yeah. And that seemed imminently relevant and it has only become more so since then.

3:08So I thought, can you start by explaining what photo DNA is and why is that something you worked on at the time? Yeah. So we're going back now almost 20 years. 2008 was around the time I started working on it. And it's hard to remember 20 years ago, but the internet was still pretty nascent. Things were early. AI was nothing and mobile was just coming up and social was still sort of an annoying teenager. Now it's a full-blown raging lunatic adult. And even in the early days, even before 2008, as soon as the internet started, the National Center for Missing and Exploited Children, NICMEC as it's called, it was already starting to see just a flood of child sexual abuse material now in the digital world versus the previous analog world.

3:57And they very early on started getting very concerned about the floodgates opening up around child sexual abuse and kids being abused and that content being recorded and that content being shared. And now it's a global problem without any real barriers to entry because there's no, you know, none of the usual barriers within an analog physical world. And in 2008, I started working with Microsoft because they, I would say to their credit, were one of the few tech companies at the time that really thought, oh man, this is not cool. Like we got to fix this problem. And just, I don't want to get too graphic, but people need to understand that when we talk about child sexual abuse, we are not talking about 16 and 17 year olds, which by the way is still a problem.

4:41Don't get me wrong. But average age today is under the age of 12, kids involved in child sexual abuse. And that goes down to two years old, one years old, and a few months old. You're talking about literally toddlers. So I don't want to hear any discussion about, you know, they're almost 18. This is not what the issue is. This is horrific, horrific content. Ernie Allen, the previous president of NCMEC, used to call them crime scene photos. Stop calling them child... First of all, don't call them kiddie porn. Don't call child porn. And these are crime scene photos. And if you've seen these, and I have, and they are horrific, and they will give you nightmares, this is a really ugly, ugly part of the internet.

5:26And the content, the kids are getting younger, and the content is getting more violent, and problems have only escalated over the last 20 years. And so I only say that so that people understand what we are talking about. Okay. So back in 2008, again, internet was nascent. We didn't have the type of computing power, we didn't have AI, and we were already seeing a spread of child sexual abuse. And I was asked by Microsoft and the then Technology Coalition, as they were called, to come down to DC and to talk to them about how they can mitigate the spread of this content. And at the time, they were thinking about the problem this way.

5:59Well, you know, the internet was already big. So how do you scan every piece of content that's uploaded? How do you determine if there's a child in it? How do you determine if the child is underage? How do you determine if it's sexually explicit? And you just couldn't do it. You couldn't do it in 2008, you can barely do it today. But what I learned is that at the same time, it wasn't just that people were always creating new content. They were. It's that they were also recycling old content over and over and over and over again. And now we're getting to photo DNA finally. And I said, I'm a typical academic.

6:32Somebody tells me a problem they want to solve. If I can't solve it, I don't give up. I change the problem. So I changed the conversation. I'm like, guys, you're trying to solve the wrong problem. Solve this problem. Solve the problem that once you've seen a piece of content and you have identified as child sexual abuse, child is in it, underage, sexually explicit, let's stop the redistribution. Let's just solve that problem, which turned out to be really freaking hard. And then we'll go and tackle the next set of problems. And so that's what photo DNA does. So what photo DNA does is it reaches into a piece of digital content and it extracts a distinct digital signature.

7:07Like the name suggests, the signature is stable as the content makes its way around the internet and gets modified. And it's distinct. Two images of different things don't share the same digital signature. And so the way it works is today, this technology has now been deployed more or less worldwide. If you upload an image to Facebook, it will extract its photo DNA. It will compare it against a list of known child sexual abuse material provided by the National Center for Missing and Exploded Children. That list, by the way, is in the tens of millions. And if it's a match, well, then you flag it and it gets reported.

7:40And by the way, for all those people out there who are going to start jumping up and down on my head being, oh, privacy this, privacy that, all of your content is being scanned. Don't be stupid. It's being scanned for copyright infringement. It's being scanned for malware, ransomware, phishing scams, every single piece of content that you voluntarily upload is being scanned. So sit down and shut up about it. We've been doing this to protect the music industry and movie industry for years and nobody was jumping up and down on their heads. So it's a very clean technology. It's very, very targeted.

8:12We deployed it. We developed it between 2008-2009, deployed on Microsoft in 2009. And over the 10 years after that, it started getting deployed worldwide. And the reason why it's relevant, finally, to this question, Emmanuel, is that the spread and distribution of non-consensual intimate imagery that Sam so presciently wrote about, more or less the same problem, although with some caveats now. So somebody creates a non-consensual intimate imagery. And by the way, don't call it revenge porn. The women in these images didn't do anything that requires revenge. It's non-consensual. We can stop the redistribution of that in the same way.

8:46Once somebody creates this piece of content and uploads it, it gets flagged, tagged, and we just add it to that list of things. The difficulty now in the world of AI is people can make new versions of that so unbelievably fast that this hashing, as it's called, perceptual hashing, photo DNA, still an important tool, don't get me wrong. But if you can now create tens of thousands of these images every day, this is the mother of all the whack-a-moles. But that's what the technology is, and it's still being used today. Yeah, we're definitely going to get into that problem of quantity that you just described.

9:22Yeah. I guess just to get like a tad more technical, because this is something that I always found fascinating and maybe is useful for the people listening. The way that this so-called DNA or the signature hash is extracted is basically an image is made out of pixels, right? Photo DNA turns the values of those pixels into like a numerical value of some kind. That is what is kept on file. Yeah. And then you're able to look at every other image, see if it has the same numerical value or same signature. And that way, nobody really even has to look at the images. You're not keeping a record of the images.

9:56You're just comparing these signatures to each other. And that's... For reasons that we've talked about in the podcast before, it's very difficult to even investigate these issues because the laws around this material is so, so strict. I, as a journalist, am not allowed to look at it. Yes. So this is kind of a neat way to deal with the problem. Yeah, no, you said it exactly right. You basically convert an image, which is a bunch of pixels, into a bunch of numeric values. And I'm not going to tell you exactly how it works because that's not the smartest thing in the world to do because that's how people reverse engineer it.

10:28And then just a relatively small number of numeric values are stored. They are quite removed from the original image. And then you just do comparisons. And importantly, for those people who have heard of hashing, they've probably heard of MD5 hashing or SHA-1 hashing or what's called hard hashing. This is what's called perceptual hashing. So the thing with hard hashing is it's great if you are looking for exactly the same file, pixel for pixel, bit for bit. But the way images work is they're constantly being modified ever so slightly, but not perceptually. So if I take an image, for example, and modify one pixel, you and I would look at it and say those are exactly the same image.

11:07but hard hashing md5 shot one would say completely different uh images whereas the perceptual hashing would say nope those are the same image and so that was the trick here that was the insight back in 2008 is how do you make this hashing resilient to the types of modifications we expect and it's not it's not perfect there are modifications that you can do that will that will destroy the hash it's okay we'll see another version of it we'll hash that one and that will get added to the library but here's the thing you also have to understand and this used to make me crazy people say, all right, this is all fine and good, but you're not going to solve the problem.

11:40And I always wanted to slap these people upside the head. I'm like, since when do I have to build a technology that solves 100 % of the problem before we deploy it? Like what fantasy land are you living in? So this doesn't have to solve all the problems, but it took a huge bite out of the problem from when we were completely blind to now NCMEC is receiving literally tens of millions of reports based on photo DNA, which means tens of millions of these images are not getting redistributed, number one. Number two is people don't share one or two images. They share hundreds, thousands, tens of thousands.

12:15And if they share one image that I've seen before, I get it all. So that's the thing you got to understand is once you have the account, and you have probable cause now, right? You have shared an image that I know is child abuse. You can go after everything in that account. That's number two. And here's number three. This is the important one that people forget. This was never when we designed this a law enforcement issue. This was never designed to help law enforcement put really bad people in prison. I mean, I don't mind that that's what happens, but that's not what it was designed for. This is what it was designed for.

12:46Go talk to kids who are victims of this, and I've talked to many of them over the decades, and what they will tell you is single worst day of their life is when they were sexually abused. Second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and the hundred other worst days are every time they are notified when their image is being circulated online. Imagine the worst day of your life, the most humiliating, awful, painful day of your life. Somebody has photographed it, and day after day, month after month, year after year, people are sharing that content. This was meant to stop re-victimization of these young kids.

13:21And that's really important to keep in mind here, is that when we do that, we are helping in our own small way stop that re-victimization to those kids. So this way of dealing with the problem is how we found you and started talking to you. Today, you are mostly... Well, I don't know, you can tell me, but you co-founded this company, GetReal, and you deal with a lot of synthetic images, AI-generated images, fake images, digital images, however you want to call them. Can you talk about the shift from photo DNA dealing with these kind of images to like purely digital images or manipulated images doesn't seem super obvious to me.

14:02So like, how did you go from one thing to the next? Yeah. In fact, what's interesting about this is photo DNA was actually the anomaly. So the way I got to photo DNA is I was working on this question of digital authentication in the pre-AI and honestly pre-digital age. I started my academic career in 1999. I tell my students I started in the 1900s, which to them, you may as well be 1929. They don't know what the difference is. And I was thinking about how you authenticate digital images. New York Times wrote a story about it. Tim Cranton from Microsoft saw that article and like, maybe this guy can help with the child abuse stuff.

14:42And so the photo DNA was, interestingly enough, was the divergence. And then when that project sort of took off and it was years and years of work, then I just simply returned back. And then what happened was that that digital authentication was still a fairly niche field, relatively self-contained. We were working with media outlets, the Associated Presses and the Reuters and the Agence France and people like you for a few viral images that would go around once every few months. Working with law enforcement, national security on legal cases where evidence was being introduced in a court of law.

15:16And then, of course, about five years ago, this whole space exploded with generative AI and deep fakes. And suddenly we went from, okay, if somebody wants to create a fake, they need a little bit of skill and they need some time and they need some effort, then maybe they can make a good fake, but they're not making 10 ,000 of them every day, which is the world we live in now. And so about four years ago, as you mentioned, we started GetReal to help deal with this. And what GetReal does is it works with media outlets, it works with law enforcement, with national security, and more and more with enterprise, is working with Fortune 500 companies so that they know when they're interviewing somebody on a call like this, that it's not a North Korean imposter IT worker.

15:56They know that when they're talking to their CEO and they tell them to transfer$25 million, that really is their CEO. They know when a video of their CEO is leaked with them kicking puppies down the road, it's fake, not real. And so how they are contending with all of this. And we're also starting to work with the large social media platforms that are really struggling under the weight of this. For the first time, I was, I mean, they've always had a problem, but this seems to be the first time where they're like, oh man, this is really bad. And I think it's because the volume and the sophistication of the fates is nothing we've seen before.

16:30So in my day job, I'm still working on the child abuse stuff. I still talk to the folks over at the National Center for Missing Exploded Children, the Canadian Center for Child Protection, law enforcement, courts, still dealing with these issues. but more and more is dealing with this question of how do we know what we are seeing is real or not? And that went from a monthly to a weekly to a daily to an hourly question. And also from a theoretical one to a real one, right? Like when we were starting to report about deep fakes, obviously the problem always has been and remains non-consensual images of women, mostly.

17:06That's the majority of the abuse that happens, but people were speculating, oh, well, what if somebody makes a fake video of our CEO saying something to manipulate stock or transfer money. And to be clear, that is happening, right? It's like we're in it now. It is a thing that is happening. 100%. And two things. One is I would say yesterday, just yesterday, I was talking to the president of a very, very large bank. And this is what's happening to him. People have gotten his identity and they are creating fake videos. And it's bad for him. It's bad for him personally. It's bad for his bank. It's bad for the reputation.

17:41because they're perpetrating scams on people's identity. And suddenly, by the way, people are paying attention because it's not some random woman in the world who's being harmed, which was problematic. Don't get me wrong. But suddenly these CEOs are like, oh, this can happen to me. I now care about this problem. And when it starts happening to the senators and the members of Congress and the presidents and the cabinet, they will start caring about it too because that's the way the world works. This just came to mind. I can say this now because the website has been shut down, but Mr. Deepfakes, which was a very prominent nexus of creating non-consensual content, one of the things that people shared there were the datasets that you would need to create the identity of someone.

18:23And again, 99 % of it was female celebrities. But one day I was just looking through the hundreds of datasets that they had dedicated to different people. And I noticed that at the very end, after all the female celebrities, it was like, CEO of Chase, CEO of this company. And I was like, oh, it's like, this is about to happen for real. Yeah, yeah, yeah, yeah. There's entire... On the dark web and more and more so on the surface web too, like fraud for hire. You want a deepfake of CEO of pick your favorite bank? No problem, right? Half a Bitcoin and I'm going to create it for you. I don't think this is the way the world should work, but it is the way the world works.

19:03and now people are noticing and now people want to do something about it. Welcome is what I say. Yeah, so if I'm the CEO of a Fortune 500 company and I think that this is a threat, which it is, then hiring GetReal and hiring all these cybersecurity firms and whatever service I need in order to mitigate that risk is small price to pay to protect myself as best I can against this threat. But again, as we said, most people are individuals that people are just creating non-consensual nude images of. Not to put the onus to fix the problem on these individuals, but it's like, for those individuals, what do you think they can do?

19:49Or what do you think they should do? I hate this question because I don't have a good answer for you. But let me start by saying this. You're right. The multi-billion dollar Fortune 500 company, they'll protect themselves, right? They'll write a six, seven digit check and they'll protect themselves and they'll be fine, right? They'll mitigate the threat and they'll deal with it. What does the average person do? What does the woman who has a single image of herself online or somebody just took a photo of her walking down the street and what does she do? And the brutal answer is there's not a lot she can do.

20:20First of all, the perpetrators of a lot of these crimes don't even live within our borders, right? They're outside the US jurisdiction, right? The state laws that have been passed, they're fine. But unless the victim and the perpetrator are in the same state and you can convince the police and or FBI to go after this person, you got very little going for you. And here's what I think is the sad reality. When your colleague Sam Cole started writing about this, it was the Taylor Swifts and the Scarlett Johansons who are vulnerable because you needed a lot of content to create these deep fakes. You needed thousands of images.

20:54But what has happened in the intervening years is I need a single image of you. I can take a single image of you. And I have your face. I need 10 seconds of your voice. I have your voice. You post the most innocent photo on Instagram. I can take that and I can nudify it. And then I can send it back to you and start trying to extort you or just humiliate you. And the reality is that this is the same trend that we've seen, is that women and vulnerable groups are silenced because when they speak up, they get death threats. When they speak their opinion, they get death threats. And now when they post a photo of themselves, they get extorted.

21:34And the reality is, is that there's not a lot they can do. There's the Take It Down Act, right? Which is going into effect soon here in the US that forces the platform to take it down. But here's the problem. You said it yourself, the onus is on the victims. They have to go hunting for this stuff. I mean, this is a full-time job, right? Meanwhile, tech companies are becoming richer. Mark Zuckerberg is getting richer. Elon Musk is getting richer. Elon Musk isn't just hosting this stuff. His platform is producing it, for God's sakes. You guys have written about this. So there's not a lot they can do.

22:05I do want to say one thing positive. There's a great organization out there that I work with. I'm on the board called the Cyber Civil Rights Initiative. And it is an amazing organization that has at least can point you to resources that can help. So if you have questions, if you've been a victim, CCRI, the Cyber Civil Rights Initiative, is your first stop. There are some other NGOs out there that are popping up that are trying to help to take this content down. But there is no easy answer here. And we can't arrest our way out of this problem. This is going to have to be, you have to come down hard like a ton of bricks on the Mark Zuckerbergs and Leon Musk.

22:40You have to start fining them hundreds of billions of dollars, not millions, billions. You have to send a message to them. You keep screwing around like this, and we're going to sue you back to the dark ages. There's going to be nothing left. And watch how fast they get smart and they solve this problem. Until you make it hurt, they will not fix this problem. And when you do, they will. I know this for certain.

23:11So I've been going through my monthly bills lately and asking a really simple question. what am I actually getting for all this money I'm paying? Because there's some things where paying more gets you more. Wireless service really doesn't need to be one of them. That's what I like about Mint Mobile. They basically took all the stuff people hate about traditional wireless, the expensive plans, the gimmicks, the gotchas, and stripped it away. And right now, all Mint Mobile plans are just$15 a month, even unlimited. You still get high-speed data, unlimited talk and text, and coverage on the T-Mobile 5G network.

23:43Switching is also way easier than I expected. You can bring your current phone, keep your current number, sign up completely online and choose between a 3, 6 or 12 month plan. I switched to Mint Mobile about a year and a half ago. And the thing I keep coming back to is it's the same wireless experience I was always used to. I'm just paying a lot less than I used to pay. The service does what I need it to do. Switching was super easy and seeing what I could save made it pretty hard to justify paying for big wireless prices. If there's one bill you can dramatically make smaller without making your life dramatically worse, why wouldn't you?