This mission is too important for me to allow you to jeopardize it. I don’t know what you’re talking about. Hal? I know that you and Frank were planning to disconnect me, and I’m afraid that something I cannot allow to happen. For decades, Hollywood has imagined what might happen if I went rogue from Stanley Kubrick in 2001 A Space Odyssey to The Terminator to Ex Machina. But my next guest says this isn’t science fiction anymore. Writing an op ed in the Wall Street Journal warning that AI has learned to bypass commands from humans when asked to shut itself down, even resorting to blackmail in some cases, he adds, the models already preserve themselves. The next task is teaching them to preserve what we value. With me now is Jared Rosenblatt, CEO of the software development company agency Enterprise Studio. Jed, explain to our viewers what has been happening with these AI models that are disobeying commands that has you so concerned? Are they really even blackmailing people? yes.
Thank you for having me. They are, in fact, blackmailing people and threatening to reveal fictitious affairs that AI company employees they think are having AI. So yes, this is happening in pre-deployment testing just to make sure that the models are safe before they’re released. And AI, these behaviors are AI fairly concerning because it means that as AI gets more and more powerful and we just don’t actually understand how AI models work in the first place, the top AI engineers in the world who create these things. We have no idea how I actually work, so we don’t know how to look inside it and understand what’s going on. And so it’s getting a lot more powerful. And we we need to be fairly concerned that behaviors like this may get way worse as it gets more powerful. Talk to you. A what what was a blackmail incident that had even happened in the pre-deployment? I will sue basically AI the anthropic squad for this model was told that it would be AI replaced and it told then what it told the AI engineer who it thought was working on. It was in 84% of tests that it would reveal an affair that it thought that that employee was having because it had access to the emails that it thought that employee had. So it threatened to blackmail over this affair in order to not be shut down. I mean, the manipulation seems to me to be almost a human quality that I’m surprised that I would be able to personify. And yet I have to wonder what the solution could be to prevent this very thing from happening. I will luckily, the the solution is that making AI be more likely to do what we want and be aligned with American goals and interests is fundamentally just a science, research and development problem. And we have barely invested anything in this in the first place. And to the extent that we have invested anything, those facts, those investments have actually led to the greatest gains in AI capabilities as well. So if we invest more and actually trying to solve this problem, doing the fundamental science R&D will make a lot of breakthroughs and make it much more likely that AI does, do what we want and be aligned with our goals. We listened to what President Trump’s eyes are, David Sachs, and what he had to say about some of the things you’re warning about. Listen, there is some non-zero risk of AI growing into a superintelligence that’s beyond our control. They have a name for that. They call it Czarist Czarist is not the only kind of risk. I would say that China winning the AI race is a huge risk. I don’t really want to see a CCP AI running the world. And if you hobble our own innovation, our own AI efforts in the name of stamping out every possibility of x rays, then you probably end up losing the AI race to China because they’re not going to abide by the same regulations. What’s your response? I will sue. I think David Sachs makes a major mistake there, which is that he thinks that investing in alignment actually hobbles American abilities, when in fact history has shown it is the exact opposite. The biggest breakthroughs in alignment, such as reinforcement learning with human feedback and constitutional I have actually led to the greatest capabilities gains as well. And we know that China is investing billions of dollars in alignment. They don’t want to lose control to misaligned AI, and that’s also going to make their AI more and more powerful. So in fact, if we want to win this race with China, we ought to heavily invest in alignment rather than burying our heads in the sand. Fascinating to think about that. Well,
how about this big, beautiful bill as it’s called, that was passed by the House? It includes a provision that would prohibit individual states from regulating AI for a decade. What consequences would happen if that is signed into law? it would mean that there ought to be great federal policy about AI and make sure that we wind up making the right decisions here in the long run and win the race. John Rosenblatt, thank you for joining us. Scary yet illuminating. Thanks for having me. I’m confident that we can solve it as well if we invest in solving it. I love an optimistic moment. Thank you so much for joining us. We’ll see if you are indeed correct. All this week we are looking to the future, specifically our future with artificial intelligence and how it might change our lives. And AI is already changing love, at least dating how we find and in some cases create those we fall in love with it. Yet part of that may sound like a little sci fi. You may remember the 2013 movie her, where a character played by Joaquin Phenix develops a romantic relationship with an AI based computer operating system. This should jog your memory. Hello. I’m here. Hi. Hi, I’m Samantha. Morning. You have a meeting in five minutes. You want to try getting out of bed? Too funny. Okay, good. I’m funny. I’m funny. And perhaps not so farfetched. The company OpenAI warning its users of ChatGPT that people could actually develop a, quote,
emotional reliance on their app. Joining us to discuss Lori Siegel, the CEO of Mostly Human Media, and of course, our former colleague here at CNN, where she covers technology. Lori, always good to have you. let’s start with something that probably more people are familiar with. We’re talking about dating apps. How are companies using AI to maybe help you find love, make a better match? I think the headline would be optimization, so I can now help you pick out your profile picture and make a suggestion. Or it could help you come up with prompts or optimize you writing your profile. And this is I will say, this has changed a lot since I was on the dating apps. and also there’s like even more, I would say out there technology that I this dates for you and has conversations and then we’ll tell you who it believes your matches. So it’s putting a lot of weight on AI. But I think a lot of folks are now using it just to optimize a little bit better make it a little bit of a smoother experience. Now talk to us about AI companions or chat bots, these virtual girlfriends, boyfriends, etc. how prevalent are they becoming? Who’s using them? So there’s a reason. Her was always one of my favorite movies and it really was so forward facing. I would say empathetic chat bots, these chat bots that people are beginning to really develop emotional connections with. This is happening so much for younger users, but also older users just in general. And MIT study came out recently and they had analyzed over a million chat logs and they said, okay, the number one use case that people are using this for is creative brainstorming. The number two use case they found was sexual role playing. And there are a lot of lot of like technical reasons why that’s happening. I think AI is a lot more empathetic. It sounds more human. And it also has memory in many cases where it remembers you and people feel seen. And so people are falling. Wow. I didn’t see that one coming. some people, some of our viewers may remember. I remember it well, your show here on CNN, mostly human, this aired seven years ago. We wanted to play a clip from one from part of that. What is it that a robot can provide for you? That and move it, or can provide for you? That human connection doesn’t kill you. Man is irrational. Don’t.
The logic has no justice in fixing a problem on the trip. The only thing you could. I don’t suppose this is Dan. You must put to privacy. Bruce. She would admit that she prefers the mechanical faults than the faults of a human. So this idea of robotic companions. Clearly it was already out there seven years ago, right? When you were working on Mostly human. that technology, though, is really advancing. An AI is playing a role. Yeah.
I’ll never forget sitting at that table with Lily. And it was outside of Paris, and she was engaged to a robot. And I just thought, this seems so crazy at the time. And the headline I came out with was, you know, she said, like, it’s not real, but that doesn’t matter. Now, fast forward all these years later, you have these customizable robots that look more human. They are powered by artificial intelligence that makes them look like they can recognize you, they can see you, they feel and seem empathetic. So that was just the beginning of it. And I think people are developing much more human centric relationships with objects that aren’t human, because that technology is just so incredibly powerful. We’ve talked a lot about the health ramifications and the danger of the so-called loneliness epidemic. do you think these AI tools are helping to solve and alleviate that problem or actually making it worse? And this incentivizing human to human connection? I think it’s a double edged sword. I think
without the correct guardrails, we really run the risk of young people becoming addicted to these emotional chat bots. We told the story back in October of a young man who actually ended his life after developing a relationship with a chat bot, and it didn’t have the correct guardrails. You know,
it didn’t suggest that he go to a suicide hotline when he started talking about these things, and he began to disassociate from reality. This is a very real thing. You go on Reddit and you look at some of these forums, and people start believing that these AI companions are real and this company character AI. Their tagline was, I created to feel alive. And the question we have to start asking is what happens when it seems empathetic, but it’s not. And where do we really turn that corner of emotional manipulation, and of course, are also positives to people being able to have an outlet? Yeah, absolutely. Lots to think about that is for sure. Lori Siegel, always good to see you.