Fireside Chat: Misinformation, Disinformation, and AI with Ruby King

I'm not gonna ask you to introduce yourself. Well, you probably should because it's a le if it's for Okay, all right. Okay, please introduce yourself. I'm Christopher Fenn. I'm the co-founder and chief data scientist of TrustInsights.ai.

We are a marketing and management consulting firm that focuses on data and AI. Okay, thank you. And so what can you tell us about how um AI is used in spreading disinformation? Okay, so this is an important distinction. Do you mean disinformation or misinformation?

You know the difference. Okay. So misinformation is the spreading of incorrect information. It is uh you know, uh the the earth is not the earth is flat, for example. But the you you're spread someone is spreading information that is is wrong.

Okay as misinformation. Disinformation is spreading information that is intentionally wrong. So uh the election was not won by a rabbit, it was run by a dog. Um the sky is red, not blue, but you are willfully creating you're creating and uh propagating information you know to be wrong for some kind of malicious purpose. So the intent is important because with misinformation, you can mean well and still spread wrong information, right?

You can there is all these people drink COVID like, oh, you know, you you inject ivermectin horse medication and you'll be fine. That's it was misinformation, it was wrong. But it was not done with a harmful intent. It was just because of the person saying it was stupid. Um, as opposed to saying you know, the vaccines are are evil, they're 5G control stuff, where you are trying to intentionally hurt other people.

So campaigns like that were conducted by hostile governments in the US and by the US to other governments um to try and essentially conduct biological warfare by proxy by harming people's trust in vaccines and getting more people killed. Okay, so intention behind it doesn't. Intention behind it does, yes. And it reclassifies it. Okay.

So it's about disinformation then where it's intentional. Right. How is AI involved in the spread of intentional disinformation? So there's a bunch of different ways. Probably the most prominent ways are the use of deep fake videos, where you can take someone's likeness, like a a well-known politician, and you essentially map their looks over an actor doing the thing, or you fully synthesize it, where it's just, you know, it's a fully synthetic thing.

So you might have a video of Kamala Harris riding a horse backwards. Uh right? A relative innocuous example. Uh or you might have, you know, Rishi Sunak uh as a as a go-go dancer. Um didn't need that image, but yes.

Those would be things where you would be creating intentionally false information and then you would propagate it to cause some form of harm. So you would say, oh, Rishi Sunak was was you know as a go-go dancer, he was high on Coke the whole night, you know, pole dancing and stuff, and none of that is true, but it's being done to intentionally harm his reputation and and cause distrust in him. Yeah. So that's one area. However, AI can be used for any amount of disinformation, and there's some more subtle nefarious ways it can be used.

So um one thing is leveraging the ability for its it to hallucinate, so you can give it essentially skewed information and say, hey, I'm going to give you the information. You write a blog post, an email, a social media post, etc. Um, to to to share this narrative and and it can do that. Yeah. Um classical AI combined with regular coding can be used to effectively create bot networks, to create the perception that a view is valid.

So this is you see this most on the network formerly known as Twitter, where you have hundreds of thousands, if not millions of bots, and they will publish disinformation, you know, oh Kamala Harris sacrificed a goat on the full moon with you know whatever. And then all this network will spread this information intentionally, trying to target different influencers and different audiences so that the narrative gets picked up and finds its way into the common population. Oh, she did that thing. No, she never did that thing. But it was used for those purposes.

You will see this with brands, with companies where people will come up with the most absurd narratives. Oh, well, this company uses radioactive metals in their frying pan. Right? It's completely untrue. But you're using generative AI to create the narratives, and then you're using classical code to propagate the narrative within your target audience.

Okay, so what kind of topics are the worst affected by this? Is it usually just political ones, or is it also like I don't know, those other things about charities and all the other things? It's anything, it's anything. So the most the the most visible examples are things like spreading vaccine misinformation, right? To to intentionally cause health harms.

Um obviously all the political stuff trying to pr to tamper with elections. Um we saw this uh in Hungary a couple of years ago, actually, where deepfake videos were being used to to tamper with the elections by saying, like, oh, you know, so this this politician's face is being saying, oh, the election date, the voting date is this date instead, and things like that. So you're you're you're doing that. But um you do see examples with brands. I mean Twitter is actually for the place formerly known as Twitter is a great place to look for garbage.

Um because you will see so many examples of just flagrantly wrong information that then actually finds its way into prominent people's um into their verbiage into their understanding of the world. There's a famous politician in the USA who's who's running for office who just repeats verbatim garbage that's on Twitter and and just thinks that this is reality, as long as he agrees with it, it's that this is reality, and that can obviously have policy influences because if he believes that the sky is red instead of blue or whatever, um that says, okay, well now we need to paint all the buildings and you know the city of the city. It's it's there's no basis, in fact. Um but you can generally classify the harms in three big buckets money and finance, law and health, right? Those are the three the most sensitive areas where if you're if you're intentionally creating disinformation, it is to cause influence in one of those areas to something as simple as saying, oh, this is the cryptocurrency to invest in, right?

Yeah, um, and you and you're scamming people, right? Um exactly. Okay. Um, or this is the the vaccine not to believe in, or this is the the the group of immigrants, they are they're the problem. Um, and then so you would use these tools in a coordinated way to change how people think about a topic.

Okay. So with sort of the book networks on the platform formerly known as Twitter, um, then is that sort of how disinformation but can become misinformation if it's people don't even believe in it. Exactly, because a a hostile actor can create the the intentionally false information, share it with people who it aligns with their worldview, and then they spread it thinking it's true. They don't know that it's false, and so they propagate incorrect information again. Again, COVID during COVID, this was one of the biggest tests of disinformation that becomes misinformation and causes literal life and death situations.

Right, yeah. So how advanced would you say the the bots are on the platform formally named Twitter in order to spread um disinformation or misinformation, I'm sorry. How do you define advanced? Well, how good are they at doing what they're intended to do, which is cause chaos. They're very good.

Even though they are very ham-handed, not particularly you know persuasive or well written to a discerning critical eye. However, one of the things that politicians in particular have done, and this is across nations, um, in the last 20 years, is they have changed politics to become identity politics where your beliefs in a political position reflect your identity. If you are if you believe you know, if you wear a certain collar hat, you're this, and these are all the values that come with this. If you if you're queer, you have all this and that because it's so much your identity, which means that it is a breeding ground for disinformation and misinformation because you are going to be more likely to share wrong information you agree with, because it's aligned with your identity, than correct information that is that is you disagree with. We are the we are the right people.

And then you're given a piece of information that seems credible, maybe it's not the best written, and you see all the support from these bots. And like, yeah, we're right. This is right. And then it becomes very difficult to change your mind because it's so embedded in your identity. Like I'm a I'm a I'm a Buddhist, this is what I believe, I'm a Jew, this is what I believe.

And if I present you alternate correct information that conflicts with your identity, you will reject it immediately. You wouldn't, you will not even consider it. So it's a so misinformation and disinformation are being used as antidotes to critical thought. Yeah, that's a good way of putting it. Um also you mentioned deep fakes earlier.

Yes. Obviously, we've seen quite a few terrible versions of deep fakes, and then there was weird Minecraft videos with Trump and Biden that were just flat out disturbing, but that was more the voiceover. But how advanced are they if someone was like really trying? If you're really trying, they are almost indistinguishable from the real thing. Believe it or not, audio is the hardiest, hardest thing to fake.

Um it may not seem like it, but uh anyone who's got any audio engineering background can look at the spectrogram of a piece of audio and go, okay, you synthesize the voice correctly, but all the background and noise is wrong. It's synthetic, it's it's manipulated. You can tell there's certain frequencies that should be there. Like we're in a room right now where there's air conditioning running. We can hear it, but it's kind of in the background, we ignore it.

However, if I was to train a model on my voice from this clip, it would also try to synthesize the air conditioning and do it wrong. And then in a spectrogram, you go, This is this is manipulated. Video is harder to detect for deepfakes because believe it or not, we're all relatively consistent, right? You have two eyes, a nose, a mouth, facial features and things. And so if you apply a vector map to your face and you apply a vector map to my face, it's relatively simple to remap the features from one to the other.

Um that's why tools like HeyGen, for example, can take an image, a single still image, understand the facial structure and mapping, and create a moving avatar from that. And it's not what is that's one of many services. Here's where here's where you start getting into ethical things with AI. Most of these tools that people are using right now in disinformation, they're very clumsy, and they are from a SaaS provider like Hey Gen. So HeyGen can go, hey, use our stuff, we're canceling your account, we're turning over your information to the authorities, right, for doing something wrong.

Okay, yeah. There's a slew of technologies that you can run on your own computer that are untraceable. And bec the one of the natures of gender of AI models is that they cannot protect us from fakes because they they do not provide any kind of authentication when they create an image, for example. It's it's just treated like a regular image. There's no metadata that says this was made by Flux One Dev or Meta AI or whatever, it's just not there.

And there are a lot of quote standards like C2PA right now, where makers like OpenAI are trying to say, yeah, that you know we're we we're assigning our images with C2PA. That's metadata that can be stripped away in the seconds. Right. I suppose that's where it comes down to like filters into verification, and how difficult it is to verify if something is true or false or if it's like just been made to look like something. Correct.

So I'm often sent DMs because by a friend who's sort of sending me artwork from Twitter. Just saying Twitter now. Yep. Um artwork from Twitter and asking, is this AI? Right.

And a lot of the time the answer is fairly simple. It's like, okay, well, he's got thirty 14 fingers. Yes, that's AI. Yeah, but it's way too smooth. And when I go to the artist, they literally pump out like 50 a day.

Right. No prizes for being able to work that one out. Right. But some of them they are a little bit like, okay, this is a popular art style, but that doesn't mean that it is AI, and it doesn't mean that it's not AI. So it can get very difficult.

But from a tech technical aspect of with stuff that's actually out there that is trying to cause harm, misinformation or disinformation, how hard is it is it to verify something like that? Um depending on the medium, it can be very difficult to identify. So if you have a state-of-the-art image generation model uh and you have the ability to remove the metadata, it can be extremely difficult to validate, like, yeah, that was that was created by AI or that was not. Especially if it's in certain formats where like there's a blurred background, there's not a lot of detail and things like that. Um text is extremely difficult to detect now if if models are good, because a lot of the so-called tells are are issues of prompting.

If your prompting is good, you're gonna work around that. Um video is still probably the least um the least stable in terms of like, yeah, you can tell something's a manipulated just because the movement in motion is wrong. Um uh again, one of the other ones is very easy to tell is manipulated is things like um music. There's certain types of music where uh you were saying the other day, uh the way that something sounds, that's not how that instrument actually works. So someone who has a trained ear would know that.

And that's where that's where the domain expertise is gonna come into play. Where you're gonna say, you you know, there's an image maybe of Joe Biden shooting a machine gun or whatever. Right? Someone who is fluent in firearms will look at that and go, that is fake. That is fake.

Not because they got Joe Biden wrong, but because the receiver on a Sigma MCX is is not like that. The receiver looks like this. And it's just tiny little details. There was one. Um I can't use that example.

Um there was one example of, oh, it was an espresso machine. And I was trying to synthesis have it uh generate AI synthesizing espresso machine, and it could not get the outlet ports of the porta filter correct. They were always just slightly off. And if you don't know espresso machines, you look at that and go, okay, that's fine. Um it's a great picture.

You look closely if you know espresso machines, like that machine would not work because the the outputs for the porta filter are just wrong. I suppose it's the same with the images of musical instruments. Exactly. Some of those, because you can get like compilations of classical music online, and for some reason they've decided to put a rotating group of AI generated images, and it's like, okay, a piano does not look like that. So the keys are not in that order.

Right. There are far too many strings on that instrument that would not play. Exactly. What are you doing? How many fingers do you need to play this instrument?

Exactly. Yeah, it's and so that's where that AI has gotten to the point now where you need discerning eyes with subject matter expertise, specifically looking for flaws that the machines don't get right yet. Yet. And in some things, it will be very difficult for the machines to get right because there's just not enough training data. How many images do you need to have of a porta filter on an espresso machine before machine before you know AI gets it correct?

Probably hundreds. So someone has to sit there and synthetically create all those and then merge that into a big production model. Now, if you were a company that made espresso machines, it would absolutely be worth your time to do that part right and just get that part and have your custom model. But your average nefarious hacker who's trying to fake a picture of you know Joe Biden riding a dinosaur is not going to spend time tuning the model to get all the little details right for a single image. Yeah, okay.

So we've covered those things, but then it comes down to the slightly more, I suppose hopeful side of it. Like how do we stop that? Or how do we even go about trying to tackle the situation? Because we've said that it's really hard to spot. Yes.

And it's rampant. Yes. It's only going to get worse in both of those. Correct. Especially as certain elections are galloping towards us.

But even aside from elections, like if we're talking about things that are more like localized governments or making the most of a crisis, how or is there any way that AI or other tools, or just humans ourselves, are able to slow this down, stop it, or challenge it somehow. We're doomed. It's not that we're doomed, it's that we have to adapt to the new reality. Okay. And that means things like I call them conduits of authenticity, where a brand or a person or a campaign has a 24-7 verification, like, hey, I just saw this image of you on Instagram.

Is it real? And you can say, no, it's not real. Um, I'm not riding a dinosaur, just to be clear. Um but and having crisis communications teams that are fluent and extremely agile. One of the things that gets a lot of people into trouble with generated content is someone generates something that's misinformation or disinformation, and the affected party doesn't respond in a timely manner.

It's like a day, a week, no comment, no comment, as opposed to two minutes after cycle, that ain't real. And and and you can shut it down right away. Um it requires so it requires a change in operations for communications, for marketing, et cetera, to say we need to have a conduit of authenticity is that anyone can call at any time, send up the bat signal, say, is this real? And we can confirm or deny it's not it is or is not real. Okay.

Um are there tools that can find these images related to a company or something for the company so that they can then say that isn't real, that is real if someone's questioning something, or is it just a case of you need to have terminally chronically online crisis PR teams ready to go all the time. There are tools, so social listening tools, a number of talk uh social listening tools that can look out for brands and logos, for example. Um Talk Walker, which is a Belgian company, uh uh no Luxembourg company, they're they're pretty well known for their visual identification. Um where it starts to get into very troubled waters is with facial identification. So if you are Rishi Sunak, right, um there are tools, a tool like ClearView AI that can uh unearth every known photo of you based on facial recognition.

In the EU, that is a prohibited use case of AI because it is so dangerous, it is so high risk. Um, and it can it it can do the technology like that is more harmful than helpful. For example, suppose you're a reporter and I'm someone who is in the witness protection program, right? And you just post uh someone in someone across the room at the another table just takes a picture of the the dining room at the at the restaurant we're eating at it identifies you, identifies me, and then you can just by facial identification reverse engineer who I am and who you are by our identities online, and now suddenly we are in a whole bunch of trouble because I'm supposed to be in witness protection, you're not supposed to be meeting with me, and so on and so forth. So from a monitoring perspective, there are tools like this, but they come at a cost.

But for like the brand identity stuff, yeah, that those tools exist and they have existed for a while. Okay, so they can be used to hunt down that kind of thing. Exactly. So the best thing to do in those situations is to talk about it, not try and not bring I think some people seem to think it's like not bringing credibility to it by not talking about something, maybe, or do you think there's another reason for them not to do it? Generally hiding your head in the sand is a poor strategy.

I mean, I I think that's true of every emergent technology, just pretending it's not there typically does not go well. Because at the moment, uh most recently, sorry it's talk politics, but there's the Taylor Swift thing where they'd done like an AI generated thing of hers promoting Trump, which was it was fairly clearly AI generated. Some of it, one of them looks fairly realistic, but it wasn't. Right. But she didn't respond to that for quite a period of time.

She only responded to that today. And well that's had a huge effect. But before that, she hadn't She hadn't responded to it. You can be pretty sure her legal team responded vigorously and privately. It does privately help though, if his fans are going, ah, Taylor Swift likes it.

And I would, and this is speculation on my part. I would speculate that her legal team's like, no, we're going to sue them for hundreds of millions of dollars first. So you just can't say anything until we get a s you know a summary judgment or something or an injunction from a court, and then we can pursue the we can pursue this with the full force of the law. Because one of the things about a deep fake like that is you are leveraging someone else's opinion uh per uh appearance and persona without their consent for commercial purposes or something, basically something other than journalistic purposes. And in doing so, at least in the USA, that is a violation of a whole bunch of laws that will cost you a lot of money.

Okay. Okay, I think that makes sense, thank you. Um so with crisis teams that uh need to be sort of taught that hiding your head in the sand is not the correct answer, and this kind of thing can snowball and have long-lasting context consequences for the for that group, and I suppose other communities too, depending on what kind of deep fake it is or just disinformation, misinformation. So they need to respond to it in a timely manner. Yes, um, and make it clear that it is not real.

Yep. Do they need to prove it's not real, or is it just a case of like it and people need to have a certain amount of trust in the group? Yeah, so generally the well, so I guess it depends there. It depends on the legal ramifications. If you're just saying, hey, that's not real, then you know you just come out and post on Instagram, whatever, say this piece of things, this piece of garbage is not real, whatever.

If, however, you are attempting to establish a chain of evidence for a lawsuit, then you're gonna want to follow a different set of procedures where you say, like, hey, this is what was done, this is when it was done, we've documented it, this is the time that it happened, stuff like that, and we're saying, you know, this is these are the remedies we're gonna seek because we're gonna sue the pants off of you. Yeah, okay. Okay, I think that's all of the questions I have. Okay. So thank you very much for your time.

Thank you. Very much appreciate it. Thank you very much. If you enjoyed this video, please hit the like button. Subscribe to my channel if you haven't already, and if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.