FMC Fast Chat: Inside Deep Fakes with Hao Li

FMC Fast Chat takes you inside the news so you can be in the know in 30 minutes. Hosted by Fair Media Council CEO and Executive Director Jaci Clement, Fast Chat features notables in news, media, and business.

You know, deepfakes have been in the news for the past few years, usually as a source of blame and often to mislead the public. But not much is actually known about deepfakes, and even less points to the positive applications of the technology. To get a handle on what all this is about, we're happy to have with us today Hao Li, who is reportedly the best deepfake artist in the world. He is CEO and co-founder of Pinscreen, a startup that builds cutting-edge AI-driven virtual avatar technologies. He is also a distinguished fellow of the Computer Vision Group at UC Berkeley. MIT even named him to their list of top 35 innovators in the world under the age of 35. We're very fortunate to have Hao joining us now.

Hi, Jaci.

How are you?

Very good. Thanks for having me on the show.

No, I'm so happy to have you with us to talk about this because it's a super important conversation to have, and I think very few people really understand it. So to have you with us is really a great public service. And I want to kick off. I know you have some things you want to show us, which is great. So if people are listening to the podcast, you'll be able to actually see the video on YouTube on the Fair Media Council channel. But before we even get started with that, I want to know, you know, how did you even get started?

Right. So yeah, to give you a little background about myself. So I work in the area of computer graphics and computer vision. And the area of computer graphics deals with, traditionally, with making all these visual effects that you see in movies, in, you know, animations, and basically creating things that don't really exist in this world and try to make them look as real as possible. So that's sort of like the field we're working in, and it's actually a subfield in computer science. And my— almost my entire career was actually, you know, started off with back in the '90s when VFX started to take off. Examples like Terminator 2, Jurassic Park. I was fascinated as a child when I saw these kind of movies and you have things that basically allow people to create anything they wanted, right? Anything that you could imagine, you could create. And so I started to get interested into these type of technologies that allows you to simulate something that looks real, but is completely artificial, right? That you can create these things on the screen. So that's sort of like how my career started. And, um, you know, later on, uh, when, you know, I did my undergrad in Germany, I did my PhD in Switzerland, I started to look into, you know, one of the problems, which is how do we create humans, right? Artificial humans that look real.

Mm-hmm.

And back then that was one of the holy grails because People could create water that looks real. People could create creatures, could create all kinds of effects. But humans were always something that was very difficult to do because we're basically social animals and we look at each other every day. We talk to each other. And one thing that's really interesting about humans is that we're very sensitive to how we look. So if someone looks a little sick today, you can tell. And the same thing happens in, um, when you create an artificial human. If there's one little mistake there, you can actually tell that this is not a real person. And that's a phenomenon that we call the uncanny valley effect. If you try to get to something, try to create a human as real as possible, if you make a little mistake, you can immediately tell that something is wrong. So that's a psychological effect. So that's really interesting because then it means that we were working on you know, something that we think is sort of like a moonshot project. Like, how do we create the perfect looking virtual human? So that led to me, you know, work on different technologies that would allow digital artists, film producers create these kind of content, you know, make them look better, make them more accessible and cheaper to produce, right? So Hollywood could basically create anything if you spend $1 million or even more.

Mm-hmm.

But we wanted to Develop technologies that allows, you know, people to create it at much lower cost and much faster. So that was sort of like the beginning of my journey into like creating virtual humans.

Okay, so how long have you been doing this now?

Well, I would say my— probably almost 20 years. So that's a long time. And Everything started, you know, when I was an undergrad. I, you know, developed, uh, you know, early versions of 3D scanners. Um, during my PhD, I developed technologies for, you know, tracking faces. Some of that technology actually ended up in the, uh, uh, Animoji technology in the iPhone X.

Mm-hmm.

And, uh, some of the work, uh, later on in still tracking faces ended up in, for example, the reenactment of Paul Walker in the movie Furious 7. Those were basically early technologies where, you know, we aimed for creating really, really faithful-looking virtual humans.

So let me ask you, what kind of grades did you get in school? Because it sounds like you have to be very smart to be able to do this.

I don't think— well, I was— I wasn't bad at school. But let me put it like this. I think nowadays you don't necessarily have to be a genius to do many things. I think one character— and that's also how we hire people and how, you know, when I was a professor, when we— how we choose students. The ones with the best scores aren't necessarily the ones that I think are critical for— I mean, it doesn't harm if you have good scores and you're smart. And if you don't have good scores, then maybe that could be a red flag, may not be, but it's an indication of, you know, how hardworking someone is, how smart someone is. But they are not the most important— they're not necessarily the most important factor. I think what's more important is if someone— I mean, these are projects that usually take a long time, right? So they take years doing a PhD. Any project that we do aren't things that you can do in a few months. They usually could take years to develop. So I think people need to be really fascinated, passionate about the problem itself, and can, you know, need to have the ability to commit themselves to, I want to be the first to do this, etc. So I think there's a lot of other factors that are almost more important than having good scores at school.

From what I understand, deepfakes have really progressed a lot within the past few years and within the past months or so. So now we're at the point where it's really hard to actually distinguish what's fake and what's real, right? Even for someone like you?

Right. So it is definitely possible to create deepfakes that with the naked eye, you won't be able to tell anything. And you can even fool an algorithm that is designed for detection to not be able to detect if something is real or not. That is something that I think cannot be avoided. And we're at a stage where you can already create like a perfect deepfake. And it's not that hard if you think about it. So, you know, when deepfakes just came out back in 2017, many of the videos that we see online, you can You can tell from the Nikita. You can see that there's some blurriness around it, something unnatural about it. But, you know, the algorithms have evolved. Also, people have understand better how to create them. So if you add a little bit of a production effort into it, meaning that, you know, you carefully select the data that you're training for, you use some algorithms that are a bit more sophisticated, you do a bit of engineering in there, And it's— and then you can always do tricks like degrading the output video, adding noise, reducing the resolution, simple tricks that basically makes it really, really hard, if not impossible, to even tell if something is a deepfake or not. And a lot of it has also to do with the performance of the person. So if you find a really good body double of someone, the only thing that you make sure is that the face replacement looks natural, then you're done.

So when you say you, you mean anyone can basically do this today?

So with the tools that are available out there, you can definitely do a really, really good job. And as I mentioned before, you can, if you have a low-res video as output, it should be enough. And, you know, one of the things that we have shown recently is that even for a high-resolution, you know, HD video, um, you can get all the details in there and still create something very, very convincing that could fool state-of-the-art, uh, deepfake detectors.

So that is kind of terrifying when you talk about it. Talk about the good applications as well as the bad applications, because, you know, when we hear about deepfakes in the news, we hear a lot of it being used, uh, in politics in particular. And we seldom hear about their application for something like healthcare. So why don't you walk us through?

Okay, so I can go a little bit over the slides and then we can see a little bit what we have. So let me first start with a fun deepfake that we've created. And let me, let me just show this. This was actually done with a comedian. His name is John DiDomenico, and he was a— he was— I think he's a really great Donald Trump impersonator. We're not sure if these kind of jokes still make sense these days, but I think he did a really good job. And one thing we wanted to see was people's reaction when they see a deepfake of Trump on him. So let me play this real quick.

Okay.

Hello everyone, it's me, President Trump, the greatest president in the history of presidents. There's some disturbing things on the internet The internet, and I know the internet. No one knows the internet like I know the internet because I'm smart. I'm really, really smart. I know technology, and I have a really big brain. And my big brain is telling me there are some fake videos out there. They call them deepfakes, deepfakes. And there's this company, Pin Screen, Pin Screen, and they've got like this genius guy running it. He's not he's not as much of a genius as I am, but he's very smart with this. Technology, this technology. So if you see me in a video and I'm not saying how incredibly smart I am and how big my brain is and how I have the greatest memory in the history of memories and how I aced my cognitive test, it's probably not me.

So this is a classic deepfake, which means that we have footage of someone And you would basically train a model from Trump video footages and build a model and do the face replacement. So what we do is we face swap the identity from the underlying actor and make it appear like the one from Donald Trump. So that's one version of deepfake. Another form of deepfake is, you know, we call it a puppet master. Which is we take an existing video of someone. So here's an example of Mark Zuckerberg from an existing video. And what we'll do is we'll basically manipulate the facial expression of him. So this was actually done by, I believe, a company called Kenny AI and through an artist called Bill Posters. Let me just show this real quick. Imagine this for a second. One man with total control of billions of people's stolen data. All their secrets, their lives, their futures.

I feel really blessed because I genuinely love the process of manipulating people online for money. Right.

Um, so this is basically, in some ways, this is really convincing because you're not putting someone's face on someone else. But the limitation though is you can, by detecting the video, its footage itself, you can say where the video came from. That's what we call provenance. If you do a reverse image search, you might be able to actually say, oh, this came actually from another source, so I can detect that something has been manipulated. There's another form which is called a cheapfake, which is you don't need any sophistication. You can just slow down a video and make someone appear differently than they are. So this is an example from Nancy Pelosi where people just slowed it down to make her make people think that she was drunk. The point is that, you know, nowadays it's pretty much possible, anything that's related to anything that's human-centric, to make people do anything they want. It can be indistinguishable from reality. And what's really dangerous here is that it can be fully accessible to people.

Yeah, especially when you get into politics, because it's just been such a hot topic. These past 4 years or so.

Right. And so regarding politics, right, so I think the first thing that is interesting is maybe a year or 2 ago, there was a lot of noise around deepfakes. I mean, the first time when deepfakes came out, the first, you know, kind of topic it was related was mostly in pornography, non-consensual pornography, right? So the first time that I heard about deepfakes was our investors calling us and saying, was this you guys? And I was like, no, it definitely wasn't. So someone, still anonymous person, implemented a deepfake technology, released it online. And one of the things is that people started to— internet trolls or hobbyists started to use this source code to put celebrities' faces onto pornography. And still, over 90% of all the deepfake content online are related to non-consensual pornography. So this is an obvious problem. When this news came out, people started to wonder, could this type of technology be even weaponized for politics, for example? And the kind of scenarios that people were thinking about are things like, we— what if we have a a leader of a country making an announcement that they're starting a nuclear war. These aren't scenarios that really happened. And there were very, very few instances where, you know, for example, some activists in, I think, Belgium made a politician say something and started to— started some viral videos, make them say wrong statements. But where this kind of technology had been used for related to politics was the following, right? So first of all, there were instances where people used something called the liar's dividend, which is they say they were caught in some scandals and they would say, well, it was a deepfake, right? So that's something that's really hard to protect against because you can always blame something to deepfakes. There were instances where political, political oppositions use the fact that you have deepfakes to say that, well, you know, the president of Gabon, you know, he might have passed away. So the new video that surged from him is a deepfake, right? So they would make, make these kinds of statements to mislead people. And something that happened even more recently is the former president making a statement the next day. which is with high likelihood to be authentic, but certain political oppositions would say, oh, it was a deepfake that was made to spread fake news.

Right.

So I think this is sort of like where this kind of the existence of this technology can be used to spark conspiracy theories, right?

Yeah, yeah. And you think that comes down to because the public really doesn't understand it?

I think the public, I think they may or may not have heard of deepfakes. I think being educated or knowing that there is something like deepfakes, I think has another issue, which is, I mean, first of all, people would question. I mean, the good side is people would question if something, whatever they see, if they can trust it or not. But I think that the second problem is they might use that fact and say, well, it was a deepfake. So that person never said this, right?

Mm-hmm.

So, you know, the fact that you always have another reason to blur facts and fake information, yeah, you know, can be— is something that I think is at the heart of what's the type of problems we have nowadays.

Right. Yeah, definitely.

Yeah. So these are a couple of examples of the potential threats that we have. I think one thing that's kind of interesting is where are— you were talking about what's coming up next, right? We have like things like deepfakes. But are things getting better? Do we have new capabilities? And what are we expecting, right?

Yeah.

And, you know, the whole, you know, deepfakes started in 2017, but this entire area of AI synthesized media pretty much started back in 2014 when deep neural networks became extremely popular for for all kinds of computer vision and AI type of problems. And they basically enable not just tasks like recognizing things or recognizing objects, recognizing faces, they can be used to generate image content. And the entire research field is very fascinated about this because there's a lot of applications that it could enable. But one thing that we have to be really careful about is what— and we have to be responsible about— is like what kind of capabilities they could create when it comes to generating content that could deceive the public, right? And so the first thing is that deepfakes, or in general AI-synthesized media technologies, are getting better. So here's an example of a, you know, an experiment we've done to generate a really, really high-resolution deepfake. So this example is, you know, Jair Bolsonaro. The guy on the right, this is a person. So this is a test we've done in collaboration with a media company in Brazil called Globo. And the person on the right is an impersonator. So they did some makeup. To his hair, to his body, to be similar to Jair Bolsonaro. And on the left, we have Jair Bolsonaro. So let me play this. And this is a high-resolution deepfake. Boa noite. Dirijo-me a todos os cidadãos brasileiros para dizer que nunca houve injustiça tão grande quanto a prisão do presidente Lula. Que crime esse homem daí cometeu? Ele não tem triplex na praia. Ele não tem sítio, coisa nenhuma. Quem nunca botou o nome dos netos no pedalinho disso daí? Se bem que eu colocaria só o nome da Raiz. So we're also very active, and for example, at the World Economic Forum, active in terms of raising awareness of deepfakes. So at the World Economic Forum exactly a year ago, we had an exhibition to world leaders where we basically show how fast this kind of technology is advancing. And one thing we wanted to show is that DeepFace can be done in real time, right? So we modified some of the algorithms that we had to basically show how we can do very, very convincing face swapping capabilities in real time. So you can see one of our engineers who's basically using only a webcam, and the webcam is tracking his face, and he can click a button and turn himself into Leo DiCaprio. Now imagine if it's real time, not only can you generate deepfakes much faster, but you can basically have a Zoom call and you might not be talking to me. Right. So this is actually another potential threat for, you know, potentially fraud.

Right.

So if we have a meeting with someone and someone is trying to make me think he is another person, that's one way of doing it.

Right.

Yeah.

You know, I half expect you to walk away and find out you're not even actually there right now.

Right.

Swear to me you're right there, Hal.

I swear.

Okay.

And the other thing is how this kind of technology becomes more accessible, right? So around the times when deepfakes came out, we actually developed the technology not for deepfakes. I mean, our purpose was to create virtual avatars for gaming and for, you know, applications and entertainment. To allow people to generate personalized avatars. But this kind of technology, if you think— if you use it in the wrong way, could basically make these kind of deepfakes more accessible, right? So here's an example where from a photo of a person we can generate facial expressions. So the AI imagines the facial expressions of that person, and you can see that it's even imagining how the teeth would actually look like.

Yeah, it's amazingly real.

Right, right. And that's just from a single photo. And here's a work from UC Berkeley where they have the ability to— well, that video is a little short, but they have basically the ability to generate a full body. So you can make people literally do whatever they want. Here's an app called FaceApp. So this app for a photo can make you age, change your hairstyle, change appearances. So you have to think about that. It's not only about swapping someone's face, making someone say something, making someone do something. You can really have more advanced controls allowing people to do anything they want. So one of the things that we started to wonder is how do we protect ourselves against deepfakes? Yeah. Big companies like Google and Facebook are releasing datasets, creating deepfake detection challenges for researchers. We have been working ourselves with— we have collaborated with UC Berkeley, with Professor Hany Farid, who's like an expert in media forensics, on how do we detect deepfakes by trying to use analyzing soft biometrics of people. So we analyze facial expressions, we analyze hours and hours of footage of people we know to match if a deepfake that is driven by someone else would match their performances. And so you can do many things. You can improve algorithms for detection. But I think the first thing you need to do is basically educate the public that there are— not only that deepfakes are possible, but To what extent are they evolving? And what are the things we can actually trust in the future? And basically form a new type of opinion about, you know, if something is likely to happen or not. Improve tools for not only detection but also characterizing if the content is So not only detect if something is fake or not, but also developing new types of technologies that could basically say if something was created for a malicious purpose. If I'm creating a deepfake for, you know, entertainment, for comedy, maybe it's not necessarily a harmful product. And of course, policies, right? So policies are regulations, basically, if you're using these kinds of things, for pornography or for harming someone, harassing someone, then this type of content should be banned. And people, you know, you should enforce certain laws for that. And, but I think policies are the most effective where they can cause the most harm. And I think big tech platforms have been quite active in I think not just focusing on deepfakes, but also on fake information in general. So I think Twitter, Facebook, YouTube, all these big companies are actually, I think in the past few years, have developed quite effective mechanisms. And there's still a lot to be done. It's not perfect yet. But I think censoring things that could cause immediate harm Et cetera. These kind of things are really important.

Yeah, well, it seems like the kind of thing that if I come up with a policy to protect people, you come up with the next generation of whatever it is to manipulate the information. Right?

Right, right. That's— well, that's true. But I would argue that, you know, you can always put so and so much effort into making something harder to detect, etc. But at some point, it may not make— it's good enough so that it makes, you know, the— if it's like, you know, a few cases out of compared to something that scales, it's already worth doing it. So for example, if you want some— so, you know, email, let's take email, for example. Who uses encryption, 64-bit encryption for emails? Nobody. But we could, we could have fully safe things, but nobody does it because, you know, it's too complicated, it takes too long, it's not worth the trouble. But if you have some level of security, some level of safety and precaution, I think it's enough to basically make sure that most of the people don't have the incentive to do it.

I see.

Okay. That is already a good enough thing. And, you know, there are many ideas that people have been thinking about, right? So what if on social media platforms, you don't necessarily have to censor things, by the way, you can even have something more sophisticated. Like let's say you have a little box that would appear and tell you that, well, you know, that information says this, but there's also facts that claim the opposite. Form your own opinion, right?

Yeah.

So because we also have to be careful about like, do we really want to censor things? Because censoring things can also become something, you know, that would cause a danger itself.

Yeah, it becomes a slippery slope. If you start censoring, then where do you stop?

Yeah, right. Which leads to the question of, do we even want to censor the technology at all, or people working on it? Well, that's, you know, almost a philosophical question. I think that is with any technology. Any technology can be used for good or bad. But one thing that's interesting about deepfakes is that when you ask someone what are deepfakes good for, maybe they might not, you know, there's— when you think about deepfakes, deepfakes is an application. The application itself, in my opinion, doesn't have many good applications. But the core technology in there is something very useful. And that has, and I believe will change the way visual information will be formed in the future. Right now we use standard 3D computer graphics, etc., but they're changing the way we display information. Let me show you a couple of examples that are interesting, right? So here's a work from OpenAI. So OpenAI is this startup company that was co-founded by Elon Musk, right? So it's an AI company that sort of like operates like an R&D lab, and they do a lot of really Cutting-edge research. And here's a work where they basically show how they can develop an AI using state-of-the-art NLP to generate images from high-level instructions. So you say, I want an armchair that looks like an avocado, so it just generates that for you. These things don't even exist.

Okay.

Now, this is really next level if you think in terms of AI synthesis. It generates the content that even for a human, it's hard to imagine. And it looks like something that you can completely— it looks like real. It looks like a real photograph. So a lot of these things are based on things that are core technologies similar to how deepfakes work. This is a work from Facebook Reality Labs, right? So how they think in the next couple of years, people will be using VR. So I mean, this is a great technology. Imagine, I mean, we're chatting here through video chat. It's all 2D. I don't really feel the presence of the other people in the same room.

Right.

Imagine if you have a technology where people wear maybe a less bulky VR headset, but can see the other person in 3D. Not only can we have a better experience for communication or teaching or education, we could collaborate. entire work would change, right? Um, so this is a game changer. It will take a while, I think, for this to become, uh, consumer accessible, but these are first things. And then the core technology here to make this thing look photoreal is almost the same thing as deepfake. It's called a variational autoencoder. It's a core technology insight that basically avoids the character that you see here to look like a video game character. Here's some other— yep.

So let me just say, if I'm interpreting you correctly, basically what you're telling me is, you know, when I live in the real world, I see things in 3D, and when I go online, I will start experiencing things in 3D. Is that where we're going?

Yes. So, okay, basically, so Right now, if we want to have a physical, like in-person meeting, right?

Mm-hmm.

We need to travel to the same physical location. If I want— my mom is in Germany. So if I wanted to talk, if I wanted to meet her in person, I have to fly there or she has to fly here. Now, all we can do now is video chat. So we see each other in 2D. So if I wanted to teleport her, into, let's say, my room here or my living room and just chat with her. Just the same thing we see in Star Wars or in other science fiction movies where you have this hologram, right? So hologram technology is a little far, far from reality, but you can imagine wearing a VR headset or an AR headset where you see 3-dimensional objects. Now you can display the other person, but I mean, if the other person looks like a video game character, nobody wants to use that technology. But what if the other person really looks like the actual person? Then to make it look real, the core technology is what deepfakes are using.

That's right.

It's something that generates something that looks like a photoreal thing and could be shown in 3D, right? So that's sort of like the direction that we're heading toward in how to make these kinds of things accessible. So I can show you a couple of things that we're doing here at Pinscreen. We're basically creating virtual humans because we believe that in the next couple of years, people are going to interact more with machines than real humans. And it's kind of a radical way of saying that, but we are already. We're talking to our phones.

Right.

We're working with our phones all day, but there's no human-like experience. And there are many things that doesn't have to— I mean, I think we like to interact with real people, of course, but there are many things where we don't have the ability to. So just think about people getting— kids getting an education right now. So they would have to do a Zoom call right now because of COVID and they can't really go to a classroom. Or many people just can't go to classrooms. They have to do watch things online. But the future is where people can actually see a person teach them.

Okay.

But you don't have so many teachers that can teach everyone. You want to have— everyone would like to ideally have their personal teacher. Not everyone can afford that. What if the teacher is an AI that knows everything? What if my doctor is my personal one, right? So there are many things that we can basically create digital twins for. And they could basically serve people. So that's sort of like the direction we're going to. And then one of the components is to make something look real and appear real and behave like a real person. That is using a very similar technology to deepfakes to create a realistic appearance. And we're working with basically with fashion companies with like Zozo Technologies, which is basically acquired by Yahoo Japan. And we're basically working with them on creating virtual influencers at scale so that brands can have their own virtual fashion models to create all kinds of fun content that we can put on Instagram and TikTok.

Got it. Let's talk a little bit about the healthcare application, because I know you had mentioned you could actually use it to share information so that your doctor is actually smarter.

Right. So let's imagine— so there are multiple things in healthcare that could be done. So first of all, let's say if we created a virtual assistant, a personal virtual assistant, we could basically have an interface so that people can get immediate support. And we actually get a lot of requests from that, from places where the population is aging. And, you know, many of their friends are, you know, have passed away. And basically, people just need to have like a companion or something.

Right.

So, or having immediate assistance, right? A nurse that you can scale and you have the— can have the ability to provide immediate assistance. We have clients who are, you know, doing COVID testing. And for COVID testing, they, for example, want to have, you know, people want to schedule a COVID test, might have questions, right? And they would like to have a Q&A Like a customer support kind of thing to also educate people. And the only way people do it right now is to have scheduled with a call center with someone that, you know, they need to wait in line. But what if you just can have 24/7 access? And it's not just a chatbot. It's an actual thing that looks like a human and people can talk to. And people would trust these avatars because you can program the avatar to have all the knowledge. So it might know even more than a real human.

Right.

I think that's sort of like where things have to go. Another thing that is, I think, even true for AI in general, doesn't even have anything to do with AI humans, is that these kind of deep learning-based algorithms are extremely powerful. So when you think about a simple example like tumor detection, right? So if now you take an X-ray, you show it to an expert, you know, a physician, and he will say, well, you know, there's like maybe 70% chance that this could be a tumor. And his performance depends on his experience and his skills. Now, it's already known that you can basically train these kinds of deep neural networks to outperform humans. And this is an example where you would not want to have a human have inconsistent results based on their experience. You basically want best possible classification of if someone— something has, you know, is a tumor or not, right? So there are many examples where these type of technologies really— these type of deep learning-based technologies make sense.

So that's the part of the story we never hear about.

Right.

One question that I keep coming back to is, if we go back to the movies like Jurassic Park, okay, where the dinosaurs look so real, how far away are we from me actually driving down the street and seeing a dinosaur cross the road?

Well, I, I think, I think if it appears in a physical space, it would be, um, I don't know how this would be technically possible, But if you wear glasses, I mean, what people understand as augmented reality these days is if you tell your— you take your phone and then you point to something and you insert some content in it, that's already possible. And I think companies like Snapchat and ByteDance are doing a great job on that. They're enriching experiences and video content that people create with these kinds of augmented content. But if we're thinking further away and if we use a bit of the You know, inspiration from Black Mirror, then, you know, maybe the question is, when are we going to have like eye implants where, you know, we mix with— I think this, I mean, could happen and might take a few years, I think, until maybe a decade until this becomes something that society would accept.

When you talk about this technology, you're quick to underscore the responsibility that you have to use it for good. And I'm sure not everyone else out there does that. So how do we make sure we keep you working for the good guys?

So first of all, we are working a multi-year contract with organizations like DARPA to advance research and development for all kinds of fake news and deepfake detection. So we're making good use of our technological developments to make sure that, you know, everyone, at least who is working on detecting if something is a deepfake or not, has access to what's possible. At the same time, I mean, I mean, even from the beginning, I'm pretty sure that many of, you know, people who are working in our field have never thought— I mean, we have never thought about like how this could become a problem for fake news. I think the ingredients for that was basically that it became more accessible. And platforms like social media basically catalyze the creation of uncurated content from anyone, right? So there is no verification about facts from any content that people generate. So that might be the issue, right? So, but this triggers like some conversation on how things can be improved. And I think— I don't think people are not doing anything. So I think there is a good development there as far as I see. But in terms of applications, you know, we're focused on like what, you know, on building technologies that are going to impact positively society and also creating tools that's, you know, that are, you know, for entertainment and for the creatives, things that are, you know, we've always been passionate to build. But at the same time, one thing that we've learned is that these technologies advance so quickly that we have to basically be really responsible. You know, we have to be responsible and be able to predict what kind of potential consequences these kind of technologies have. The Internet Media Council is a 501 nonprofit organization. Advocating for quality news and working to create a media-savvy society. For more information about the Fair Media Council and upcoming Fast Chat shows, check out fairmediacouncil.org.