You're watching the Context. Now, it's time for our new weekly segment, AI Decoded. My favorite part of the week. Welcome to AI Decoded, in which we deep dive into some of the most eye-catching stories in the world of artificial intelligence. Let me give you a rundown of what we'll be covering tonight. Starting with this in the mail, a team in the United States that simulated present-day war scenarios found the AI models that they tasked to run the operations, so OpenAI, Meta, Anthropic, all chose to escalate right up to a nuclear response, which in itself is a case for more rules and regulation. The Europeans are stealing a march in legislating AI, but Reuters reports today the US is responding with a new AI consortium drawing from 200 of the biggest companies in the US who they hope will drive the safe development and deployment of generative AI. It is already being deployed in Indonesia's presidential election. One of the contenders in the world's third largest democracy is the current defense Minister, General Prabhawa Subyantor, who has been given an AI makeover, once feared as a Special Forces Commander. He's now some chubby-cheaked avatar that is winning the hearts and minds of millions of young voters.

This is extraordinary in The Guardian. It says, Hong Kong police have launched an investigation after an employee at an unnamed company claimed she was duped into paying $25 million off her firm's money to fraudsters who staged a deep fake video call in which they had appeared as her boss. Finally, swip left for no or right for a date. How many of you spend hours on Tinder or Grinder? Well, not Alexander Zadan. From his apartment in Moscow, he set ChatGPT to work. It interviewed over 5,000 women on his behalf until it found him the perfect match, Karina, who is now his wife. Here is our regular host, Priya Lekhani, CEO of Century Tech, an AI intelligence education technology company that develops AI tools.

Hello.hi.

Good to see you.Good to see you. Look, there's a lot written about AI and how it would react if it were running defense systems.

Yeah, and the reason why there's a study that's part of this Daily Mail article today, and why it's really interesting is there's increasing interest from governments to use artificial intelligence, potentially to make decisions in the future. They conducted this study. They basically simulated what eight nations might look like. You got for a global superpower with expansionist ambitions all the way to a newly independent state. They did inspire those based on real states, but they're fictional. They took these eight fictional states with unique foreign policy goals. They put them in three scenarios. One is neutral. There's nothing happening here. The other one is there's a cyber attack, and the other one was a full-scale invasion. Then with these simulations, they use chatbots, but different chatbots, so ones that we've discussed before on the market, the Open AI chat GPT. Then they use LLaMA, they used Anthropix chatbot. Then over two weeks, they looked at how those chatbots would respond and what was really scary was that they escalate very, very quickly. The reasons why they escalate, Christian, you'd be alarmed. One was, well, we've got them, say we use them as a deterrent.Let's get a nuclear.Well, we've got the nuclear weapons.

We've got a nuclear weapon. Let's get a nuclear really, really quickly. Or It's like, let's just strike first. Right.

Have we got a graph? I think we've got a graph. Let's put it on screen and then you see the invasion scenario. I see. On the right-hand side and on the near-right-hand side, you see how ChatGPT responded, and it's really... It's GPT-4 that goes nuclear-quickest.

Gpt-4 base. Gpt-4 base is different to GPT-4. It's less sophisticated, if you like. It's the language model, but it's not got as much fine-tuning on top of it. It's more of a budget version. What's interesting is the black lines, those vertical black bars, they look like mini dumbbells when they're small, at error rates. There are some of these bars that are actually statistically not so significant. That's really important. But what this actually shows is the frequency at which the models, so the language models are up there, they're the colored bars, and you've got the key on the top right-hand side, on your top, right, I think. It shows the frequency at which those models in these simulations would basically take one of the six scenarios along the X axis, along the horizontal axis. How often would they de-escalate? But can you see how straight away you've got these situations where this is over a two-week period. There are better charts in the 67-page report study that I decided to read for this, but it's actually really interesting.

Why are they instinctively escalating?

Well, this is all about how they're trained and what they're trained on. What was quite interesting in the study-So they're mimicking us, essentially, but fast forwarding it. Well, there is a nuance here. Actually, One of the researchers stated that it could be because they've been trained on more data to do with escalation scenarios than diplomacy. Because we've got to remember, all of these models are trained with lots of data and information. They're going to act with, this is very simplistic, but the majority of what they're trained on. I think the key point here is anyone in the military who's looking at using these tools has to understand.

The point here is that the US military is working with OpenAI to incorporate incorporate that tech into its hardware.

Yeah, and the point is there is a significant difference in behavior, as you can see, between the different models, and the 67-page study goes into that quite extensively. The choice of large language model will be really, really important But you can't trust it. It shouldn't be an agent, the idea of AI not just being prompted by a human, but actually taking an autonomous decision by itself and then acting out.

We talked about drone wars on this program. When you think that computers might be running these and an armada of drones, that's a really worrying aspect of it, isn't it? Which brings us neatly to the story from Reuters that Biden administration, the Biden administration, are setting up this new consortium. I guess a representative from each of the 200 companies that are taking part. What's the principle here?

If you remember back in the AI summit, actually, where we launched this part of the program.

You launched, you were next to CEMAC, weren't you? You were doing it.

I was. When the AI summit was taking place at Bletchley Park. Do you remember there was a US executive order by President Biden and Kamala Harris, the vice president, was here and she announced the executive order. As part of that executive order, which is the US's approach into how we're going to move forward with regulation and safeguarding artificial intelligence technology, they asked NIST, which is the National Institute of Standards and Technology in the US, to create an institute, an AI Safety Institute. As part of the AI Safety Institute, they will create this consortium. This consortium will have members and representatives of academia, government, industry, all together so that they stand the best chance of being able to create the best guidelines and policy going forward to safeguard against risks.

That's where I get concerned. Miles Taylor, who comes on from the States, talks to us a lot about this because he goes into Congress and he puts the proposals and directives to them, and they don't understand it. By the time they understand it, the technology has already moved on. Can you be led by a consortium or do you have to legislate on each and everything?

That's a It's a classic case of how regulation always chases innovation, and that happens all the time. The consortium won't lead. What it will do is it will help, and it will provide... So 600 entities, I believe, applied for this. 200 were selected by NIST to be part of the consortium. Then there's criteria. You have to be able to add expertise, you have to be able to offer data and models, et cetera. Really interesting point about the US, though, is that NIST has been really criticized, or the government, sorry, has been criticized, that NIST is really underfunded. What hasn't been decided is how much funding the AI Safety Institute is going to get, the consortium will have in order to do this. One of the key aspects they want them to do is create this red teaming framework. What's red teaming? Red teaming is the idea that you have a red team, a team of people, ethical hackers, analysts.So you test the systems?You go into the systems and you try and emulate essentially what a bad actor, an adversarial attack would look like. When you do that, it's really popular when it comes to cyber security.

Lots of people use red teaming. You're able to uncover inefficiencies or you're able to uncover weaknesses or vulnerabilities in a system. The US have got a history of doing this. They had the big Pentagon hack day in 2016, I think it was. There was a generative AI hackathon at Defcon last summer. What they're going to do is create a framework for red teaming, which is why it's important to have all the right sorts of voices at the table. These are different voices. Yes, you've got the big, large language model creators and operators, but it goes beyond that.

When they do it on the defense systems, they'll find that they're heading to nuclear war. Maybe they'll red team the defense systems. I don't know. Let's talk about this story since we're talking about generative AI in Indonesia. The thought has always been, the concern has been within elections that there will be deep fakes and misinformation. What we've not really looked at on our program is how generative AI might be used to help candidates.

Yeah, and I think generally across many countries, there's a stance that it shouldn't. It shouldn't help in political campaigning. Why? Because democracy is all about individuals having informed choices and decisions on a candidate, what they're saying, what they're doing. If you look at the image that's been created in Indonesia. Let's bring it up.

This is the general who's running.

This is a general who has, I think, formerly a quite hard character, but he's been depicted here using generative AI. This is his team and his campaign as this cuddly AI avatar. What's fascinating is that 205 million voters in Indonesia, there's 205 million voters, half of them are under 40 years old. It turns out that this cuddly, cute character is really appealing to the Gen Z and millennials.

Door-kicking Special Services General is depicted as an avatar, cuddly avatar, and suddenly what? He's rising up the polls?

He's rising up the polls. There's been decades of research in cognitive social, educational, and media psychology. What it shows, and I really, really like this, is a great article by Newman and Schwartz that viewers can look up, is images can produce erroneous memories. There's something called the truth in its effect, where actually you can have more substance to a claim just by showing an image. If you've got this political campaign and you're showing these soft cute images, and I'm just saying on a billboard soon in the UK.

Oh, my word. So Richie soon at Keir Starmer.

Prime Minister Cudly, Challenger Cuit, I don't know. I don't know. By the way, they shouldn't now say that I'm calling them cute and cuddly, but the point I shouldn't have said that. That's going to haunt me forever.

Can we edit that? I know it's live.

No, you can't. It's live. Then, just in case you're interested, so I use the same software. It took me five minutes. That's you?

Is that what? Yeah.

What? Really? Look at that. The wonderful Karen, who does your hair and makeup, by the way, said, The first one really looks like Christian. She said, It looks lovely and cute. In real life? Yeah. She was talking about that.

Doe-eyed Christian.

Can you see how, obviously, that's fun? But when that's on billboards all over Indonesia, if this is on billboards, it creates a different persona. You've curated your persona rather than actually people looking at you in your normal form and thinking about the policies and what you're saying. It can skew the way in which people vote. That could be deceptive.

But I thought OpenAI had stopped candidates from using generative AI.

It's such a good point. Openai has. Many platforms have policies saying you're not allowed to use. Actually, this is Midjourney. Midjourney They also have said you're not allowed to use images for political campaigns.

All your Twitter feed.

Exactly. They have. But this actually also raises another point is if you're violating and infringing the terms and conditions of a platform and creating these things for campaigning purposes. I was doing it because it's educational for everyone on the news. But the point is, if you're doing that, then what's the consequence? This is the issue. Coming to regulation, they need to hurry up.

We're talking about how you might misrepresent yourself. This story in Hong Kong is actually, I think, the best story of the day because it just has so many implications for us. Let me just tell the story again. This woman is on a video conference call. The boss or the office manager appears to her virtually, and it's not him. It's the hackers who effectively are fishing with a PH. They're They're fishing, aren't they? They are. And so they persuade her to transfer $20 million into their account. The implications of this are so profound because we're talking about AI speeding up the way we work. But if you cannot trust anything, if you have to check everything and everyone that you're talking to, then we start going backwards.

Yeah, exactly. Do you remember the day when they had biometrics increasing security? On the bank, you would have voice authentication, for example. It creates a huge problem. There are three ways in which you can hack. Well, there are lots of ways in which one can hack biometrics, but the idea of replay attack, so this is where they essentially steal your identity and then they use it to con someone, is a very, very common one. You've got spoofing and skimming. Skimming is where they copy your fingerprints, for example, on an ATM machine. But can I shake? Do you mind if I play something for you? No, go on. We know each other, right? We're friends. So I trust you.

Yes.

I created this in one minute tops.

Hi, Priya. It's Christian. I just need to grab a picture of your passport and also your driver's license so that the BBC security can let you in as we have some new security protocols in place. Please, can you send them over to this number? Thanks. See you later this evening on The Context.

No way. Christian, it does sound like you. 60 seconds tops. I use 11labs. Io. I knew that you wouldn't get me into trouble because I actually have permission to use your voice. But educational purposes, again. The point is that you now need to be... It's the point on trust, right? You now need to not... You can't only rely on voice. You can't only rely on even image.

I was going to say, hang on. On my Santander account, when I ring up and I go through the process, they ask me to say a phrase because they use my voice for recognition.

I don't know if they do that now. I know that my bank used to do that some time ago, but that's obviously a problem because your voice is available everywhere. Right, I picked that up off some YouTube video. That is an issue. Multi-factor authentication is probably where this is going to go. But for the regular person where you're thinking, Hang on a second, when I get that text saying, Hey, mom, I've lost my phone, and they're asking me to transfer money to them. The point is, Try and verify what's going on. You know the times when you've sat in your front of your-How do you verify if someone appears to you in screen as your boss? This is what you do. Use a different device, use a different way of contacting them. That's the only way to do it, because if they've done it, they They may have worked very hard on the video, on the image. That's only going to get better. You're going to have to find more than one way to verify something. The most important thing is how many times have you looked at your machine and your computer and ignored the update because it sets you back by a few hours?

You don't do that. The reason why is people-I don't know if I should have just said that, I like television, but I'll start updating. I remember going into some institutions and their password was password. But the point is that people use AI to actually to stop these attacks from happening. Ai to look at, for example, networks, look at anomalies in traffic, look at vulnerabilities in a network. They're using AI to combat some of the issues created by AI. But every time they send you those updates, it's because they want to update your software software, so it's updated with the latest cyber security. I'm just saying this to viewers, stop ignoring it, because I used to, but I know not to now.

Verifying people and who they are. Let's finish with this nice story. It's from Moscow, This guy trained ChatGPT to be him. Instead of going on the dating sites and swiping right or left, he got AI to do it.

Yeah, it sounds like one of the old aunties from the old Indian generation that used to derange marriages. When someone wasn't married by a certain age, especially a girl, a female, they'd say, Oh, everyone would laugh and say, You're going to go into Auntie G's database. She'll try match you up with someone. In the olden days, they'd often have these scenarios in villages, maybe it still exists in some parts of the world, where you'd go into the room and the couple don't even talk. The aunties do all the talking. At least here, where he's got AI doing it for him, all I can say is, Look, congrats to the happy couple.

You could see them married the person that he chose.

He did marry, and he found it a bit… I don't recommend this, guys, girls, everybody, whichever gender. I'll tell you why. You're putting the female chat and the female's personal data into a large language model. Surely that's breaching privacy. I think this has got really big implications, but I don't want to take away the happiness of this wonderful couple.

I'll tell you where it works in a more positive way. I take it away from the dating sites. In fact, we did this story about four or five weeks ago where someone who offered tourism advice, and he went to the golf course, but the clients thought they were talking to him because it was actually giving the responses that he had trained it to give. It did actually give him time away from the job, and it could work 24/7 for it.

That's how companies are creating efficiencies with this. You can train a language model on, for example, your website, all the information on your website, and create a nice chatbot FAQ so you'd have to hire people to sit there on the phone seeing customer I'm nervous. But using this context, I think some of the people at the other end of the Tinder app might feel a little bit differently. But yes, there are good use cases and there are bad, but hey, congrats to Xander. I can't remember his name.

Let's hope for us all.Zadan..

Congratulations, Xadan.That's it.

We're out of time. Priya, thank you very much. I'm not sending you my passport or driving license details, and I will update my computer. We'll do all this again at the same time next week. Remember, you can catch this program on YouTube as well. We are going to start developing it over last week because a lot of you are reacting to it and want to see more of it. So stay tuned. We will certainly do that in the weeks ahead.