Can We Trust AI? Safety, Ethics, and the Future of Technology

Stephen Horn and Autria Godfrey ask whether artificial intelligence can be trusted and whether it is safe. Shyam Sundar, director of the Center for Socially Responsible AI at Penn State, discusses why people trust conversational AI and the mental shortcuts behind that trust. Duncan Eddy, executive director of the Stanford Center for AI Safety, explains what AI can learn from aviation safety and how researchers try to verify AI systems.

Watch on YouTube

Transcript

Hello and welcome back to Agents of Tech, where we explore the minds of the greatest innovators addressing the world's critical science and technology challenges today. I'm Stephen Horn, joining you from London.

And I'm Autria Godfrey, coming to you from Washington DC. And today we are looking at AI. And we're asking the question, can we trust it? Is it safe? AI technologies are becoming integrated into almost every aspect of our society, ranging from health care and finance to transportation and education. As technology continues to evolve at a rapid pace, questions surrounding the ethical implications of AI algorithms and their potential biases, and the trustworthiness of AI driven systems become increasingly pertinent. Coming up on today's show, we talk to Dr. Shyam Sundar, Director of Center for Socially Responsible AI at Penn State University, about making AI socially accountable and ethically aligned with our values. And we hear from Dr. Duncan Eddy, an expert in space satellite communication and now the Executive Director of the Stanford Center for AI Safety.

That's a big job, isn't it, Laila?

Yeah. Massive job. But we're really happy, as usual, to be joined by Laila. Great to see you.

Good to see you, too.

What are we talking about today?

So we're talking about AI safety, and how users interact with AI, as well as more of a technical detail of building safe AI. It's a vital topic because the rate at which AI is accelerating is kind of exponential. And we're starting to build this resource and almost infinite supply of intelligence, which we didn't have. And we kind of need to figure out how we regulate it and how we distribute it fairly.

It's truly transformational, right? The number of things that AI can do. But I don't know about you, but I get the feeling that a lot of people just don't trust it.

There's this argument that the faster that AI develops and the more advanced the technology gets, the safer it gets. So there's also that side of it, and I think it'll be interesting to explore that in the conversations with Dr. Eddy and Dr. Sundar, does advancement correlate with safety?

So Autria, where are you in all this then? Do you trust AI?

Every new technology, you have to take it with a grain of salt as it's evolving and we're kind of working out the kinks. One of the things that I think is really interesting is, you know, the Nobel Prize winner and who some referred to as the Godfather of AI, Geoffrey Hinton. You know, he's really kind of expressed his skepticism that AI still has some issues and that, there could still be some concerns about AI and whether or not it can be fully understood. So I think that, you know, it's interesting that you have him calling for more research into AI safety and, you know, saying, look, this is something I created, but I still think that we need to have safety parameters in place to make sure that the information that is being, you know, regurgitated and spit back out in these AI models is the truth and is something that people can put stock in.

But the trouble is that how many of us actually understand what's going on, you know, we can ask ChatGPT what we want, but no, none of us actually understand what's going on inside the box, if you like. So it makes trust an almost impossible issue, doesn't it? We have these huge corporations spending hundreds of billions of dollars, and we just have to take it on trust that they have trust at the core of what they do.

Yeah. And, you know, OpenAI recently published a paper in October where they're now using machine learning AI agents to do the machine learning engineering. Right. And they are using that as a kind of safety, test to see, okay, is it capable of engineering itself? And if it is, then it's a snowball effect, isn't it? Yeah. So, you know, if the AI is able to do the machine learning engineering, what chance do we have?

Well, all I'll say to both of you is I have seen this movie and it doesn't end well.

Well, fortunately, guys, we do have an expert with us today who can help boil all this down. Dr. Shyam Sundar, Director of Center for Socially Responsible AI at Penn State University, is an expert in how technology affects the way we communicate and perceive information. He researches how features of digital media like chat bots, social media and AI interfaces influence our trust and engagement. Dr. Sundar, our thanks to you today for being with us.

Thank you for having me. It's a great pleasure.

Let me start. So the thing we've been talking about in the studio today is AI and trust. You know, and we've been wondering about how much people trust AI. And I was going to start from your perspective and from your work. Obviously we interact with AI, ChatGPT or other large language models in a different way to we would do a Google search. You know, it's much more conversational. Do you think that affects trust?

Yes, absolutely. I think, the conversationality is an important part of our interactions with generative AI technologies like ChatGPT, like Claude and several others that have popped up in the last couple of years. And what they do is they give you the illusion of a highly individual response when they chat with you, so they take into account your specific questions and provide an answer that seems very tailor made to exactly what your concern is.

And so it seems very credible right off the bat, the fact that they are tailoring that information for you and that back and forth conversation lends it that aura of credibility.

Dr. Sundar, I want to bounce off of that and talk a little bit more, you mentioned about that personalization. And when it comes to technology and differences culturally and technology, obviously when it comes to social media, different countries prefer different types of social media, such as WhatsApp versus Facebook. How do you think that impacts AI usage and the experiences of users from different cultures? And have you done any research on how people in different countries, or from different cultures interact with AI differently?

So in general, our research is not looking at cross-cultural comparisons as much, but what we find endearing across, participants in, different parts of the world. When we do studies or when we see other peoples studies as well, is this very strong belief in machines being right.

What we call a machine heuristic. It's a mental shortcut that we have that humans in general have across cultures that seems to suggest that machines tend to be objective, accurate. They do not gossip behind our back like humans do. So we attribute these stereotypical notions of objectivity to machines, which translate to our trust in them. And so this is something that we found across cultures. When we've done studies in different countries or when other people have done studies that show that, this kind of, blind trust, if you will.

Given all of your research into AI, the trust, you know, whether we can trust or not, you personally, how much trust do you put in AI.

In general, I look at AI as not so much as an information resource, but more as an entertainment or as a starter for brainstorming, or as a provider of template these days with generative AI. I think the mistake that a lot of people make is they think that AI just because it's available in the same venue - on a computer - to a website, they tend to think of it as a sort of search engine. And I think they tend to equate it to Google and Wikipedia, psychologically. But, the way that these AI tools generate, output or content is not so much based on research. It's based on a guess, a probabilistic estimate, of how likely the next word or the next set of words are likely to be true given, these sort of seed words at the beginning. So it's just guessing what comes next. It's not really looking up information and providing you, reasoned, you know, judgment, or an answer to your specific question. So I always am leery of AI results. I always, verify that with a Google search engine or some other search tool, just to verify if, in fact, the information is factual.

I think that's really interesting. You know, when you look at, you know, computers and technology as a whole, AI is is still so new, it's still in its infancy. Do you think that we will ever get to a point where you do have 100% trust in the accuracy of what AI is churning out?

We probably will not have 100% accuracy, but one of the things that we are seeing is, a fast developing move toward combining AI with, search and other research tools, as a backend. So in a sense, it's, likely to give more of a human voice, what we were saying earlier about conversationality to what were otherwise, very mechanical, kind of interactions that we would have with machines. So you're already seeing this with Gemini on Google is a search engine, on top of which there's AI built into it so that the search engine results are communicated in a kind of conversational way, just like a smart speaker does. Like Alexa talks to you and gives you the weather. Alexa gets it from an information resource, but gives it to you in a natural language exchange, as in a conversation.

And as, AI models kind of do get more advanced and there's claims of PhD level intelligence or ability, in terms of knowledge, how do we build in kind of structures so that we can verify that knowledge which the AI is spitting out, because inevitably we're going to get to a point where we're asking it questions where the user doesn't actually have the knowledge to verify it. So what systems are we building to help users, verify information better that they receive from AI as the knowledge gets more complicated?

Yeah. I mean, in general, we want experts to be in charge of information. So for example, you can go to an AI tool to get medical advice. You can feed in your symptoms and you might get a fairly credible sounding diagnosis for your symptoms. But, it's always a good idea to go to an expert, which is, in this case, a doctor who's trained in medicine to be able to take that symptoms and even perhaps your AIs output into account, before they render a judgment. Because oftentimes the context is not, specified to an AI. AI may not see you holistically as a person. It might just see you based on, you know, the kind of symptoms that you have. And so you do need experts. But to our domain we are talking about we do need human trained experts, especially, like you're saying, PhD level, to be able to, you know, just verify, if what the AI is providing is true or not.

Dr. Sundar, I want to ask you about a very fascinating theory that you've created in the interactive media effects or your TIME theory and how that can really play a role when it applies to human and AI interaction, particularly in the context of trust and information. Can you kind of give us a little bit of background about that TIME theory and how it applies to AI and trust?

TIME basically stands for Theory of Interactive Media Effects. And, it essentially is designed to understand how the medium, especially interactive nature of the medium, is making people trust the content of the medium. So when you think about it, you know, how interactive a technology is should not make any difference to how much we trust the content that is produced in the technology. But, studies have shown over and over again that people are kind of bowled over by the interactive aspects of the technology.

And so what my theory does is essentially break that down into what we call an affordance or action possibility. And interactivity is one such action possibility where people, when they see that a website is interactive, they automatically attribute certain characteristics to it. Like if a politician's website is interactive, they think that the politician is much more welcoming of input, is much more likely to be inclined to listen to their constituents and so forth. And therefore they have positive impressions of the candidate, even though they never might engage with that interactive tool.

So that's what we call the cue route of the Theory of Interactive Media Effect, where it's based on these mental shortcuts or what we call heuristics. So for example, if you see, a story, a news story that is presented in a visual form versus a textual form, it triggers what we call a realism heuristic, which is an immediate mental shortcut that says, if I can see it, then I believe it.

Or, for example, if you go to a social media site, you might see that a particular post has been, forwarded a number of times. The number of retweets is even published on the interface. And so you say, oh, this has a lot of retweets. That means you know, it's a very important post. A lot of people think it's credible. And this is what we call a bandwagon heuristic. So these kinds of interface indicators can directly lead us to form trust based on our application of these simple mental shortcuts that we call cognitive heuristics.

But then the theory contrasts this kind of decision making with an action route, which is much more, systematic and not as heuristic. So action route is based on actual interaction and engagement. You know, when you see something, in different modalities, you actually, extend your perceptual bandwidth, if you will, to understand what it has, and then you kind of get into it. Or if you see something that is interactive, you actually interact with it and engage with it and go back and forth like you do sometimes with ChatGPT. And that, you know, increases the amount of agency you have in the interaction. And it also promotes self-determination. And all of that is a much more reasoned approach toward forming trust.

I think it's so interesting how, obviously in some domains that can be a negative, but in others, such as psychiatry, it's quite critical for the user to trust maybe an AI tool which is supporting a psychiatrist, and to feel that they have a space to express themselves, positively.

So yeah, I mean, sometimes the trust is warranted and that's completely fine. And by certain people it can be, you know, very useful. If you are a trained psychiatrist using this to better articulate, something to your patient, then it's certainly a very good use. But if you're a layperson who is using it, but, you're just bowled over by the fact that it seems to be, actively listening to you and seems to provide exactly the kind of, tips that you want, then you might be misled by it. And so that's what we call unwarranted trust. And there there's a great need for trust calibration.

Couple of questions to wrap up, Dr. Sundar, the first one is you know, I've spent my life in television, as has Autria. One of the things that we've learned over the years is how easy it is to fake sincerity. It's quite easy to see that the more interactive you make the tool, the more people are going to trust it. But that doesn't make the information that's being shared from that tool any safer. But, you know, we live in a world. And let's take that model of psychiatry. You know, we live in a world where there are not enough psychiatrists. And so there won't be usually somebody to check those results. So as we spend billions and billions of dollars on, you know, these AI models moving forward, how optimistic are you about the future? How are you optimistic, are you that as we develop these models, we're also going to be able to make sure trust is at the center of what we do?

Yeah, I think that's really what a lot of us are fighting for. And that's why we have a Center for Socially Responsible AI that's why we do a lot of research on trust. Trust ultimately is a psychological entity.

It's not so much, a property of the medium or the technology. It's based more on how people interpret information and how they consider that information. So given that I think, different kinds of sources have different, levels of credibility in the eyes of the public. And increasingly, AI, is gaining some credibility. But, you know, the moment it goes wrong, then people develop algorithmic aversion, which is, you know, the direct opposite of machine heuristic. They just become very averse to anything that AI does.

And so it's in the interest of AI companies for them to be, trustworthy in the long run. So I think having, a concerted effort not just by industry but by watchdogs like our center and other people who look out for the welfare of the users, I think we will arrive at a solution that is best for everyone.

There is right now a coalition of more than 60 tech companies, and industry groups like Amazon, Google, Meta, even OpenAI that are trying to get congressional leaders to pass legislation to permanently authorize the US Artificial Intelligence Safety Institute, hopefully by the end of the year. How beneficial would that be to your cause?

I think it would be definitely beneficial to have any kind of oversight, and any type of institutional, or governance type, oversight that would perform audits on a routine basis on these systems, where they do curation of the training data that goes into these models and also push for disclosure and transparency of the characteristics of the training data mechanisms in place, where we can systematically address some of the baked in biases and, you know, mistakes, that AI could make, in order to improve trustworthy systems in the future.

Well, thank you ever so much indeed for joining us today. We really, really appreciate your, insight. And thank you again.

So Autria, did that answer your question?

I think it sounded like we're at least, taking the steps to try to make AI, as trustworthy as possible. But it kind of sounds like until some of those safeguards are put in place, AI continues to operate as kind of the wild, wild West with no, no rules, no regulations. And so as long as that's the case, I think people should be skeptical with what it is AI churns out.

But I think Dr. Sundar's point on that is it's in the interest of, you know, those who are advancing AI for it to be aligned in terms of user safety and user trust, that, you know, the success of AI is reliant on how the user utilizes it and how the user perceives.

AI is going to and is generating transformation like we have never seen before. And the problem with that is that transformation has winners and losers. You know, there are people who will do very badly out of AI, and there are people who will do very well out of AI. And so at the moment we're trying to say, this is going to be good for everyone, and this is going to create a society which is going to be a better society because of AI. And I don't think that the trust is there for that. I don't think people today trust, you know, that type of set up. And I think that's a challenge.

No, I agree, I think there needs to be transparency about that winner and loser kind of fate in the future and kind of contingency for what is to come for people who are going to be at a loss, which I don't think we're talking about enough or nor are we putting enough effort in to build those safety nets.

I think that we just had Dr. Sundar's, you know, thoughts on it. And it's a great starting point. You know, take the information that AI churns out, but then go and verify it on your own. But our next guest Dr. Eddy should be able to help us with this. Shouldn't he as well in Stanford?

Yes.

He has spent years working in the realm of space satellite communications, And now he's directing his talents toward AI as the Executive Director of the Stanford Center for AI Safety, we'll speak with Dr. Duncan Eddy. It is our pleasure to welcome Dr. Duncan Eddy. Thanks for your time today.

Thank you very much, Autria. It's a pleasure to be here.

Okay. I want to get started by kind of going back. If we could talk a little about your background and kind of explore how your expertise from aerospace engineering can now provide unique perspectives on AI safety.

Aviation is one of the safest ways to travel today. For miles traveled, aviation is the safest mode of transportation by far, though it didn't always start that way. There is a long history in the field of applying safety validation techniques to develop ever safer ways of traveling and getting around, and these techniques work all across the design spectrum, from early design to flight testing to operational procedures. Once systems are deployed and these types of learnings that we've developed to make air travel safe can also be applied to AI systems to make them safe for wide adoption in society as well.

Can you elaborate on those systems? Just a little bit? I mean, when it comes to the checks and balances and the rigorous testing and validation, and I would assume also the redundancy that has to be in place in order to make air travel as safe as possible. How can those types of systems be, be beneficial when it comes to AI systems?

So one of the things that people usually think of in safety in terms of a system is either safe or it isn't. Aviation did not become safe overnight. My thesis advisor, Mykel Kochenderfer one of his main contributions is the development of the next generation of aircraft collision avoidance system. That system was developed after an accident in the 70s where there was a mid-air collision over Germany. Unfortunately, that resulted in the loss of both flights.

But since then, millions and millions of dollars in large scale efforts have developed systems that solve this problem of just mid-air aircraft collision avoidance. Safety isn't just a one off. A system is safe or not. It's really an ecosystem of techniques and validation where incrementally you build and develop safety capabilities and really look at any individual failure. That's the role the FAA plays, where if you have an aircraft incident, really of any kind, and they will come in, they will take your flight data and then they will inspect, incredibly thoroughly to see if you are following procedures. Or is this a new gap where there is a new type of failure that hasn't been caught, which means that then we need to develop new techniques, new procedures, new safety verification methods to build that into the safety ecosystem.

And that same type of rigor and discipline can be brought to other fields where again, it's, I think, premature to think that AI could be entirely safe out of the box. And if it's not safe, then we just throw it out. That hinders progress and the good that it can do elsewhere. What instead we can do is develop incremental techniques where we find certain types of failures. And then there's been some work that's part of the work that the center does where we develop specific mitigation strategies, mitigation techniques, where we can then start to apply those in an incremental fashion to develop a safe system overall.

Dr. Eddy thank you for that. We've been talking today about, AI safety, and I've been playing a bit of the role of the skeptic. I don't think I'm out of order here. I think if you walk around the streets of any town here or where you are and ask people, you know, their confidence in AI systems, you know, I don't think people I think there's a certain degree of skepticism about how safe those systems are. Can you talk us through a little bit about what's changed? I mean, how has the verification and testing of AI systems changed from other, say, software systems in the past?

I think that skepticism is very healthy. In terms of, how it differs. I would even say it doesn't differ too much from past systems in terms of software verification. A lot of the techniques we're using to validate AI today in terms of our so-called formal verification methods, were developed for software systems in the 60s, the 70s, the 80s, and applied today.

These techniques are essentially they go back to your grade school mathematical proofs, where you can make formal guarantees and rigorous proofs. That means you can make statements of this condition is always true. We can apply those same techniques in AI fields and that same type of reasoning. It's hard and challenging, because some of these models are very large. But we can start to develop systems and techniques where we can start to make guarantees. And then the question is, where is it worth investing to really verify these types of systems? Because it doesn't come without a cost.

I just wanted to jump in on, the point of obviously the verification methods of AI systems don't differ too much from software, but how about the rate of development and how closely they're tied? So the rate of development for AI is far faster than that of previous software or aviation, and is the safety research keeping up with that rate of acceleration?

I would say that the safety research is, frankly, trying to keep up.

AI isn't really just one thing. It's a collection of mathematical techniques that build on statistics going back decades and decades. And there have been a couple of breakthroughs in the last five, ten years that have led types of learning based problem solving to be very effective at solving problems we haven't seen before, which is then led them to being adopted widely in tons and tons of fields. That is great. We can solve new problems that we haven't previously.

It also does present new challenges, as you mentioned, and that spurs a need for more research in terms of safety, in terms of validation, in terms of verification, which they go hand in hand. And generally the safety tends to lag a little bit with innovation. But I have high confidence. Some of the highest attended classes here at Stanford are on safety and verification of AI systems.

I want to ask you about some of the different approaches to improving safety. The next generation is going to be tasked with taking this on, which is probably a good thing, because it seems like with the creation of AI and the way that it's going to be constantly changing and constantly updating itself, the regulations and the safety parameters are going to have to be updating and changing right along with it. So I guess my question is, are we constantly moving the goalposts when it comes to trying to make sure that AI is as safe as possible?

There will always be that challenge. I mean, there is always been innovation, and there always needs to become an ongoing conversation about what the risks and what the right mitigations are. We've seen a number of different attempts at regulation, from the EU AI Safety Act to SB 1047, in California. And they take different tacks from either regulating applications to regulating the development itself. I would generally be a little more skeptical, especially now from regulating the development itself, because you're essentially trying to regulate mathematics. There are also convincing arguments for large models, foundation models that only very capital intensive companies can build, that those could come with some additional scrutiny, and that could be reasonable depending on what their capabilities are.

I also think a very relevant conversation to have is how these systems should be deployed. The things I really like in the EU AI Safety Act in terms of don't deploy AI systems in college admissions and HR hiring that a lot of these systems, those are high risk applications because it's relatively easy to integrate biases when you're training them without even intending to do so. That's a large part of the research that the center does as well in terms of developing, essentially assessment frameworks for if you're building a system, what are the checklists that you should think of that you should consider when you're building these types of systems, to make sure that you don't run afoul or, you know, have things go wrong even if you aren't intending them to do so.

Dr. Eddy, I want to from that go on to the Reluplex algorithm them that your team has developed for verifying the properties of deep neural networks. Could you explain, firstly how that works? And then the real world kind of applications for that?

So it was actually developed before my time at the center though with colleagues from my time. So it's developed by, Kyle Julian, Guy Katz Mykel Kochenderfer, Clark Barrett and David Dill. And it combines the field of formal verification that we discussed earlier in terms of software, with verifying properties of neural networks. So a neural network at the end of the day is an equation that you can write down. And formal verification is a field of mathematics that deals with making provable statements about equations. You can combine those two fields where essentially what they do is they transform ReLU-based neural networks into a, Boolean satisfiability theory problem. There's a whole field of math that people have developed over a very long time to prove statements about these types of equations. And if you can express your problem, in this case, a ReLU-based neural network as a Boolean satisfiability problem, you can then make guarantees about it. For example, you can say if my input is in the range of 0 to 1, my output will always be in the range of 0 to 1.

This becomes exceedingly relevant, for what are called single pixel attacks. So you will see in terms of image recognition, where you might have seen incidents where you have a stop sign, someone slaps a sticker on it, tags it with some graffiti, and it's no longer a stop sign, it's a yield. And when deploying autonomous systems, having them recognize things in the real world correctly is incredibly important. If stop signs are no longer stop signs, that is a safety risk.

But it's incredibly hard to guarantee that, okay, my image classification system will always properly classify stop signs, even if they have weird stickers on them all over the place. Very easy problem for humans. Hard for computers. Robots can help with making guarantees about small, You can't use them on LLMs but if you have a reasonable size network, you can make guarantees of: if my input is always in this region, even if it's perturbed by daylight or rain or clouds, it will always be guaranteed to have the correct output. But that's how the technique works. And it can be used as a tool to verify small to medium sized neural networks to make guarantees about their behavior in different scenarios.

So how could Reluplex be used in other kinds of lateral fields, such as medicine. Right. So maybe I imagine for avoiding false positives, would that be possible? Or in other lateral fields such as finance.

The challenge becomes that it all depends on how you're posing the problem and what the exact problem is. But if you have a problem where you have some set of inputs, you, use a neural network to take and process those inputs and generate sets of outputs. You know, because this is the real world, that all of your inputs aren't going to be nice, pretty, human, curated, very clean data that there might be noise to it. And if you're having a trading algorithm that, you know, for example, trading stocks on the stock market and suddenly you get one bad bit of data that says, you know, Amazon shares are trading at $1,000 a share. Now sell your position. You don't want to have one bit of bad data suddenly cause your algorithm to run away. You can then start to make guarantees that, okay, even if I have one perturbation into the system, it will behave in a predictable way.

Are there any breakthroughs or kind of positive directions, promising directions that you've seen in terms of the scalability of, these formal verifications?

In terms of formal verification, I would say not so much in terms of other techniques, there is something that I'm incredibly excited for that we've been developing recently in the Center for AI Safety called adaptive stress testing, where adaptive stress testing actually uses a AI based reinforcement learning agent to probe an AI model for likely failures.

One of the big challenges of testing AI systems especially LLMs is their complexity. You can throw garbage at them and you can get them to also throw garbage, say garbage back, and jailbreak them, and lots of things. But that isn't the main risk of deploying a lot of these systems. One of the big risks is in regular conversation like we're having now. Are there sets of words where you can trigger significant toxicity out of a model? So adaptive stress testing actually leverages AIs generative power in a way to explore those types of decision spaces to find the likely causes of failure. And we've deployed it to great success in autonomous vehicles. And they've also deployed it in terms of testing large language models to figure out how you can elicit toxic responses from models that developers have otherwise thought were safe.

Just the amount of money that is being spent on developing AI capabilities. AI models by some of the largest, most successful companies in the world. Are we spending enough money and are we dedicating enough resources to safety so that we can clearly demonstrate that this is at the heart of everything that we do?

I am hopeful for the future about the conversations that are being had, but I don't think it's something where we can relax and say companies will be responsible on their own. It needs to be built into the culture of how we develop these models and other ones. Safety isn't a one or the other part of building systems. It goes hand in hand, you build the system and you also think of the safety implications. So both need to be invested in simultaneously.

What do you wish other people in the field knew more about? Like what do you wish that they knew that they don't really talk about necessarily?

I think there's a lot of noise and there's a lot of hype around AI and what the systems can do. I think the thing that people - it isn't even actually really a safety thing - it's focus on solving real problems for people and make sure that there is a core problem where the sit the AI based system is the right solution. Technique to solve a meaningful problem.

Well, thank you ever so much indeed for joining us today. We really loved our discussion and we really appreciate it. So thank you.

So Autria and Stephen, are we hopeful for the future of AI safety after that discussion?

I see people like Dr. Eddy and I'm full of confidence. And you know, he talks about culture, right? He talks about a culture of safety. Do we see a culture of safety that is at the heart of our culture? That would be my question that I'm yet to resolve.

There are going to be mistakes made along the way. There will be some collateral damage. You know, there might be some situations where AI does end up being beneficial to bad actors. But I think what gives me hope at least, Stephen, in speaking with Dr. Eddy and and speaking with Dr. Sundar is, the amount of attention that safety mechanisms surrounding AI is already getting, like AI is not ubiquitous yet, but already there is such a great focus on trying to keep it as safe as possible, which there was not. If you think back to, you know, the advent of the internet and the World Wide Web, like it was just kind of a free for all until we started realizing some of the detriment that could happen. So that's what gives me hope is like we're kind of starting out from a place of caution and concern.

I think, AI will totally transform our society and will transform it in a fraction of the speed of any other major revolution we've ever seen will do. And I am yet to be convinced that advancement for all is at the core of this. But I see good people doing good work. Right. So I'm with you, Autria, in that I'm hoping for the future because we see good people doing good work and the younger generation are there to kind of take this forward.

What I'm fearful of is, although Dr. Eddy has reassured us about safety and development of advanced AI, goes hand in hand, for some reason, it feels like advanced AI is just going to be here in a blink of an eye, whereas the safety might lag behind.

Now, Laila, what's coming up next time?

So next time we're looking at deepfakes and cybersecurity. And that will be with Dr. Siwei Lyu from the University of Buffalo's Forensics Lab, and also Dr. Damon McCoy, who is the Director of NYU's Center for Cyber Security.

Well, I look forward to that.

So do I. That's going to be fascinating. Can't wait.

But that's a wrap for today's episode. So don't forget to like and subscribe so you can get all the very latest from Agents of Tech. Until then, goodbye everybody.

Agents of Tech is brought to you by WebsEdge. Special thanks to our editor, Abbie Harries, graphics by Matthew Coleman and Agents of Tech is produced by Cath Sheehan.