The Munk Debates Podcast - Be it Resolved, Chat GPT will cause more harm than good.

Episode Date: March 22, 2023

Seemingly overnight, Chat GPT has exploded out of the deepest corners of the emerging tech space and into the mainstream, capturing the imaginations of everyone from students to CEOs. But with any new exciting technology, there tend to be more questions than answers. For the creators of Chat GPT, tech writers, and other AI evangelists, this is a Sputnik level moment in tech, and will have far ranging and transformational consequences for the future.  The impact of this revolutionary technology is already being felt, and this is truly just the beginning.  This does not mean that Chat GPT will transform the world for the better, but will without a doubt come to define life in the 21st century. But for other computer scientists, AI specialists, and the generally unimpressed, Chap GPT is nothing more than a clever party trick.  Chat GPT is not even close to artificial general intelligence, but merely a finely tuned and at times impressive mimic.  Chat GPT is also rife with errors, and is difficult to trust. A program that produces such inconsistent results is far more likely to be a flash in the pan than a technological revolution   Arguing for the motion is Gary Marcus, Emeritus Professor of Psychology and Neural Science at NYU and a leading voice in artificial intelligence. He is the author of five books, including, The Algebraic Mind, Kluge, The Birth of the Mind, and the New York Times Bestseller Guitar Zero. His most recent book, Rebooting AI, with Ernest Davis, is one of Forbes’s 7 Must Read Books in AI.     Arguing against the motion is Jeremy Kahn, Senior Writer focused on artificial intelligence at Fortune Magazine Speaker Quotes  GARY MARCUS: “I think we have to be realistic that the number of different ways in which these systems could cause harm is quite large and that some of the specific harms are quite serious”.   JEREMY KAHN: “People might have made similar arguments about the printing press and other technologies, about broadcast technologies when they came along, that these things would somehow obliterate the truth. But it actually expanded the potential of people to express themselves.”   The host of the Munk Debates is Rudyard Griffiths - @rudyardg.     Tweet your comments about this episode to @munkdebate or comment on our Facebook page https://www.facebook.com/munkdebates/   To sign up for a weekly email reminder for this podcast, send an email to [e-mail].     To support civil and substantive debate on the big questions of the day, consider becoming a Munk Member at https://munkdebates.com/membership Members receive access to our 10+ year library of great debates in HD video, a free Munk Debates book, newsletter and ticketing privileges at our live events. This podcast is a project of the Munk Debates, a Canadian charitable organization dedicated to fostering civil and substantive public dialogue - https://munkdebates.com/   Senior Producer: Jacob Lewis Editor: Adam Karch  Become a Munk Donor ($50 annually) to get 72-hour advanced access to the full length editions of Friday Focus and Munk Dialogues. Go to www.munkdebates.com to sign up. Hosted on Acast. See acast.com/privacy for more information.

When you're a journalist and people don't trust you, it's always your fault. These people need to be represented. They are Canadian. They deserve to have a voice and a seat at the table. It is time to go back to the office and the time is now. Russia had reasons to be concerned. They had reasons to be fearful. We're at an absolute turning point in reproduction. This is the problem with realism. They just treat all countries the same. They don't distinguish between dictatorships and democracy. Welcome to the monk debates. Every episode we provide you. with a civil and substantive debate on the big issue of the day to arm you, the listener with

enough information, to make up your own mind. Today's debate, be a resolve. ChatGPT will do more harm than good. It's only been a few months since we started hearing about ChatGPT, and the AI chatbot is already finding its way into classrooms. Tonight we're taking a closer look at a new technology that's making waves in the world of AI. Chat GPT, a language model created by OpenAI, has the ability to respond to prompts in a human-like manner. Does chat GPT represent a breakthrough that will spawn new businesses, or is it more of a gimmick? Is chat GPT coming for your job? Hello, I'm your moderator, Rudyard Griffiths. Well, seemingly overnight. Chat GPT, the large language learning model, has exploded out of the internet onto our screens and computers and wowed us with its ability to

write press releases poetry do complex computer coding many people are saying the arrival of chat gpt heralds a new era of machine learning that will transform work society for the better making many boring tasks redundant for humans passing them over to computers and machines this is a technological revolution that's going to sweep aside many of our long held assumptions about how we work, play, and stay informed about the world around us. A lot of AI analysts say it's as revolutionary as the internet, but some say it's a threat to society, but one thing is for sure, it's sparking interests among everyone from top CEOs to students.

But some prominent scientists, AI specialists, and others in the know are less than impressed with chat GPT. They say there has been a rush to judgment, assuming that chat GPT, is in fact some form of intelligence, where as in reality, it's anything but this is more clever party trick than anything approaching artificial general intelligence. In the few months that we've all spent with chat GPT, it's shown that it's rife with errors, hard, if not impossible to trust in terms of many of the information that it provides you. And overall, it is not the catalyst for a technological IT revolution. that its proponents would insist that it is.

On this installment of the bunk debates, we're going to explore the future of machine learning by debating the motion, be it resolved, chat GPT will end up doing more harm than good. Arguing for the motion is Gary Marcus, Emeritus Professor of Psychology and Neural Science at NYU, and a leading voice in artificial intelligence. He's the author of five best-selling books,

including the algebraic mind, and the New York Times bestseller, Guitar Zero. His most recent publication, Rebooting AI with Ernest Davis, is one of Forbes seven most important books to read on AI. Arguing against the motion is Jeremy Kahn, senior writer at Fortune Magazine, who focuses on artificial intelligence. Gary, Jeremy, welcome to the Monk Debates.

Thanks for having us. Yeah, thanks, Patrick. Let's dive right into this debate. our resolution today, be it resolved, chat, GPT, large language learning models writ large, are going to do more harm than good. Gary, you're on the pro side of this debate. So take us away with your opening statement. I'm going to start by saying that the intellectually honest answers we don't know, that there's a ton that we don't know about how these systems are going to be used, both on the positive and the negative side. And so it's hard to tally up the ledger. But I'm going to

make the case that it's certainly going to do a lot of harm these kinds of models and that we should really be worried about it. I will start with misinformation, the deliberate construction of misinformation. Bad actors are going to take these tools. They're going to make misinformation at scale. That misinformation is going to be authoritative. It's going to sound true. They're going to be able to make as much as they want for essentially no cost. And that's going to cause a deluge of propaganda like we've never seen before. And that is going to undermine trust everywhere. and I think undermining trust is bad for democracy and good for authoritarian. And so I think that's pretty bad.

Beyond that, there's a whole periphery of cases that we don't even understand yet. Every day I see a new way in which these systems might be misused or cause harm. There's just an endless periphery of ways in which these tools can be used for harm. And yes, of course, there are some benefits. And I'm sure Jeremy will make a nice case for them. But I think we have to be realistic that the number of different ways in which these systems could cause harm is quite low. large and that some of the specific harms are quite serious.

Thank you, Gary, for that opening statement. You're listening to our debate today, be it resolved. Chat, GPT, will do more harm than good. We've now got the other side of the debate. Jeremy Kahn, your time now to give us an argument in favor of large language learning models. Well, I want to start by saying that I agree with much of what Gary has just said. I think the harms he has outlined are very much real, and it is possible.

that some of these harms will get worse over time. That said, I think that overall, you know, taken in comparison to the benefits of this technology, that there's a good chance that some of the harms, which in the near term may grow over the long term, we will find ways to diminish and mitigate these harms. And as the technology improves, essentially the benefits, which are already great, and I will outline some of those in a minute, will continue to grow while the harms will essentially diminish so that over the long term, these technologies will probably have more benefit than they will harm to society. And in terms of some of those benefits, we already seen how chat GDP and some of the related systems are being used by lots and lots of people

to do things that are very time-saving. Those who we were familiar with chat GPD, it can do lots and lots of different things for you from composing letters to compose. you know, recipes to, you know, doing writing code. And while none of what it produces is perfect, lots of people have found it to be time-saving. And, you know, even if they have to go over and check the work that the system has done, just having that initial first draft of something completed in virtually no time at all is incredibly powerful. And if you look at some of the big companies that are using this, it ranges from sort of Instacart and, you know, is using it to help recommend things for your weekly shop,

and then it puts those things automatically into your cart. You have Shopify. It's doing something very similar. Snap has a chatbot you can talk to, as well as trying to snap to your friends, and users are already finding that useful. And then on a more sort of industrial use cases, you have this thing called Codex,

which is based on one of the underlying large language models that is similar to what power is chat GPT, also from OpenAI. That's being used to write code. And so I think there's a lot of, lot of potential uses of chat GPT that are going to determine tremendous benefit. And I think these harms, while real, we are going to find ways to mitigate them. And over time, we will find that they diminish. Thank you, Jeremy. You're listening to our debate today. Be at Resolve chat,

GPT will do more harm than good. Okay, Gary, your opportunity now for a rebuttal responding to Jeremy's opening statement. I completely agree with Jeremy that there are lots of positive applications. But I think it's interesting to compare codex and programmers' experience with, say, J.P. Morgan and Citibank's experience. So some people are already using it and using it well. And some people have already decided to ban it. So J.P. Morgan decided to ban it. And probably it's a temporary ban, but they decided to ban it city bank did as well.

I think Apple banned an email program that used it. So programmers are in a good position to use chat, GPT, and the like, or codex, because they're custom to debugging things. And the output that you get is not reliable and it's not trustworthy. And the average person doesn't really understand that. Programmers for a living work with code that doesn't work. Maybe you're typing something, you forget a parentheses or you forget the name of a variable. I don't know how many listeners have coded, but a lot of coding is trying stuff out. It doesn't quite work. You test it. You fix it. And when you have an unreliable assistant, which is what codex or chat is, if you are accustomed to a cycle in which you can fix things,

then you're not phased by it and you can deal with it. My understanding, though, is like only maybe 30% of the coding is right. A lot of it is not right. If you put the same stuff in the hands of non-professionals and do kind of no-code programming, which is something else that Jeremy talked about, you may actually wind up with pretty serious problems. So people who are not trained in the ways of coding don't understand how things can break. They don't know how to build unit tests to make sure that they're working,

and they may wind up with things that actually cause a lot of problems. Sometimes serious economic problems may be worse. And so it's not in that sense of one-size-fits-all technology. So, yeah, there will be some applications. The real question, obviously, and we can't fully answer it here, is about the net effect. My view is most of what Jeremy talked about is saving time, and saving time is obviously important. I think about how I use my time all the time, and we all do, and those benefits are real.

But to my mind, at least in the short term, we don't actually have any mitigations around the misinformation problem that I started with. In the United States, I'm really worried about the 2024 elections. I cannot imagine that these things will be a small part of the election. I think they're going to be a big part of the election. I think that all parties will use them, lots of outside nation states will use them to try to influence the election. We've seen that sort of thing already in the last several years. This now presents a new tool that costs almost nothing to use and creates incredibly

authoritative bullshit. That's going to be a problem. If we lose democracy because people don't know what to trust anymore, in my mind, that outweighs whatever benefits we have of programmers writing faster and people being able to write letters and recommendation faster. Thank you, Gary. Okay, similar opportunity here for a rebuttal from Jeremy. So Jeremy, let me put a couple of minutes on the show clock and turn the program over to you. Again, I probably agree with much of what Gary said. and I think where we disagree may just be on what the ultimate impact of some of this will be, I do agree that if we lose democracy over this, that would not be worth any of the benefits we might get from better efficiency

in the way a lot of us conduct our professional lives. But I don't think that that's actually what's going to happen. While I do think there's a danger here of political misinformation, I also think that it won't be that much of a difference in degree from the problem we already have with political misinformation, which, as we know, is a serious problem. But this, at the moment, most of this misinformation that's been problematic is created by people. And it's, you know, and we still have a problem with it. So I don't see it as a huge, you know, I don't see it as a difference of degree that creates a difference in kind. It's just a slight,

you know, we might have slightly more of this. And I actually think that the awareness of how easy it is to create misinformation and some of the erosion of trust might actually be a good thing, that we might find that people are, you know, more questioning of the information that they read. And that would not be a bad thing. If people are skeptical and they say, well, where is this fact actually come from? And then, you know, they have to go and find a reliable source or they have to go do a traditional Internet search to discover this information. That would not be a bad thing.

And I think, you know, people made similar claims about, you know, the advent of photography that would, you know, supercharge propaganda. And to some extent, it did allow people to create more compelling propaganda. And yet, you know, democracy did survive the advent of photography. And I think democracy will also survive the advent of chat GPT. Thank you for that rebuttal. Let's go back and forth here, gentlemen. This is terrific to see this debate unfolding.

So I'm going to come right over to you, Gary, as sense you want to get a word in here. Yeah, this is a pretty cool debate because I think we're actually rapidly narrowing down on what we think the key questions are. We actually agree on a lot, including I think that the consequences for misinformation culture are maybe the most important ones and that estimating what those are likely to be actually is central to this debate. So I think we agree on the positive consequences. We agree that if the misinformation situation were bad enough, that that would actually offset the positive

consequences. We agree on all of that. And then it's a question of, is there actually a more serious problem before and how much more serious? We also agree there's already a misinformation problem. There has been for decades or centuries, arguably longer. And so the question is like, why am I more worried right now? And it's all about volume and scale. So the Mueller report in the U.S. about the 2016 election said that Russian troll farms were spending over a million dollars a month, and they're very sophisticated in their ways of generating misinformation. They have lots of servers and fake accounts and so forth matters. We're making it possible to make misinformation that is cheaper and better at the same time and a lot faster. So,

It used to be that you had to write your misinformation by hand. Not everybody can do this. Not everybody can do this in a language that isn't their native language. Now you can be a non-native speaker and still make up lots of well-formed English stuff, and you can make really vast amounts of it. I've published in my substack some examples. Even with chat TPT and its guardrails, you can get around those guardrails. And you can make up, for example, as much misinformation as you want about something like

the January 6th invasions of the Capitol or, COVID-causing, you know, fake problems and things like that. And you can make it detailed. You can make it with fake studies, fake numbers, and essentially at no cost. And so my concern is that it's sort of like saying, well, knives have always killed people. We don't need to worry about submachine guns. Well, submachine guns make a difference because they can mass produce destruction in a way that is much more difficult with a single knife.

We have these tools that can mass produce misinformation in a way that we've simply never dealt with before. Okay, let's give an opportunity for one more rebuttal here. I'm enjoying the back and forth. Take us away for a final rebuttal. Then I'll join with some questions that are top of mind for our audience tuning into this debate. I would say, Gary, that yes, I think this is a vital question, and we just don't know exactly, you know, how big an effect this will have. But I am hopeful that the effect will not be as big as you fear. And that, you know, again, I think previous technologies that people have worried about in terms of misinformation have had some impact, but they have not managed to destroy democracy, and I don't think this will either.

And like I said, I think there's a chance here that actually people will become more skeptical of what they read, and that would not necessarily be a bad thing. I also think that this technology or the underlying models, there is the potential to use them, again, not on their own, because as Gary says, it is true. They have no understanding of fact or fiction. they don't know what a fact is. They're just stringing together a bunch of statistically probable letters to create a response. But there is enough sort of natural language processing ability from these underlying models

that they can actually be used to create automated fact-checking systems. And you may find that this technology becomes somewhat self-policing and that some of the mitigations are using the very same technology that is causing the problems that Gary is most worried about. Hey, Monk podcast listeners, I wanted to let you know about our other weekly audio program. It's called Friday Focus. And hey, guess what? It comes out each and every Friday.

It's half an hour long. And it provides you with a masterclass on international events, all the big issues and ideas shaping our world. We've got that for you each and every Friday here at the Monk Debates. Simply access via our website, triple W monkdebates.com. Click on Friday Focus in the top right navigation. You'll get all the details or check out a sample of the program in the same podcast feed as the main Monk Debates podcast. I hope you'll join us for the next edition of the Friday Focus podcast. Now back to our program.

You're listening to our debate today, be it resolved chat GPT, i.e. large language learning models will do more harm than good. Okay, Gary, let me come to you with some questions that I think of. surface for our listeners tuning into this this fascinating debate. And I think we all understand, and maybe you'd like to expound though a bit more on it, your argument as to why these models, in a sense, or anything but, quote, artificial intelligence, quote, close quote. You think actually they're kind of stupid machines. But at the same time, you seem to be saying that this time is different, that their effects are going to be outsized, they're going to be large, they're going to be profound. Is there an inherent contradiction there, Gary, between your view that these things aren't

really intelligent in any way, shape, or form, yet you also think they're going to have this profound and continuing impact? How do you reconcile those two things? You don't have to actually be that smart to cause a lot of damage. We had a recent U.S. president that, you know, I found to be fairly destructive and was maybe not the brightest among, I won't name any names. You know, The real issue is actually how much power we assigned to machines. My concern is that people, ordinary human beings, are not well trained in the art of understanding artificial intelligence and how it works and what its limitations are. You certainly don't want, for example, to hook them up to an air traffic control system

or to a driverless car. People have had them play chess and they make up their own rules. Like, they'll follow the rules and then they'll have a bishop jump over a rook and You can't do that in real chess. It worries me that systems that don't really understand the rules get treated as if they do, and we start to assign them more power. And so aside from the misinformation stuff that we were talking about, in general, the overapplication of these systems bears risks.

Like, you know, engineers always think about things in terms of, like, what are the risk case? I build a bridge. What are the circumstances in which the bridge might break down? We don't really have a clear understanding of these systems in terms of the, formal properties in which they might break down. And people have this intuition that they're smarter than they really are. Fascinating. So, Jeremy, isn't that the inherent problem here? It's not so much these large language learning models, chat, GPT, Bing, and others. It's us. It's in a sense how they fool us. Because as you say, they're just stringing together

statistical probabilities in terms of, you know, likely answers and responses. And we rather like those probabilities. We rather like the answers that they provide, but we're fallible. We've proven that over and over and over again. So why isn't this technology especially dangerous because it's so good at hacking us? Yeah, well, I think that is an issue with this technology. People are too trusting of it and too credulous of, and that is an issue. And part of it is that the systems are good enough at imitating us that, you know, it's easy to forget sometimes that you're not speaking to, you know, a human being. But I think, again, that problem may not persist for very long, because I think we're all going to become accustomed to the fact that the entity, you know,

that we are communicating with on the other end of the line might just be a piece of software and may have no actual intelligence at all. But, Jeremy, how will you know that? How will you know? Like, unless there's a good actor, unless there's a well-intentioned actor, but what if it's Gary's scenario of a Russian troll farm. Yeah, yeah. No, I agree that there may be problems where people do not identify that they are putting the software out into the world to interact with people. And that, you know, is not ethical.

And I think, although they don't exist right now, there's plenty of discussion about creating, first of all, some standards that would apply, but perhaps also some laws that would have some real teeth where any, certainly any legal entity, any company or government that wants to put these systems out there would have to disclose to people that they are interacting with a piece of AI software that they are not interacting with a person.

That does not solve the Russian troll farm problem. It does, you know, if a criminal wants to set up a bot and use it to convince you that, you know, it's your aunt sending you your text messages and she's in trouble and needs to send you money, it's not, you know, that's not going to solve that problem. But we already have laws that say it's illegal to con someone into something.

something like that. And, you know, we'll just expand those laws to this new technology. And I think in the meantime, you know, we will have some new laws come in that will handle, you know, entities that are regulated and any company that wants to put out a product where they will have to tell people that they are interacting with an AI system and not with a person. Well, let's have Gary come back on that. Because Gary, you know, Jeremy brings up a good point. You know, technology is iterative. And already, I understand, professors, universities are running chat GPT essays through what are so-called AI. I don't really like that word, but a device that will test whether it's written by

a human or not. So aren't we already kind of solving in the ways that Jeremy thinks that we can to allow us to have agency and control over this technology as opposed to being controlled by it? I guess there's a couple of things to say there. One is those existing tools are actually pretty lousy. So none of them work better than like 80% at distinguishing a human versus a machine. And if you've written something and there's a one in five chance that somebody's going to say, you just have to give up what you said because we all think it's a machine and that algorithm is wrong.

You're going to be pretty frustrated. Like those things would have to work extremely well to be useful. And 80% is just not going to cut it. Second thing to say is the regulation question is really interesting. I found myself in a little fight on Twitter yesterday with Dominic Cummings. Nobody on my side of the pond will know who he is, but I think a lot of people on yours will. It's pretty interesting that Boris Johnson's advisor was getting into this, and he was arguing against regulation. And this is a sign, I think, of things to come.

There's going to be a lot of arguments against regulation. One view is going to be that all the existing laws cover things. We don't need anything else. In the case of misinformation, at least in the U.S. laws, which I know are a little bit different than in the UK, you can pretty much as much as you want without consequence as long as it's not, for example, defamation or something like that. We don't really have laws set up to deal with the mass production of misinformation, where you might have a troll farm that produces a million or a billion pieces of fake information each day.

You might want to treat that kind of wholesale production of misinformation differently than the kind of retail or smaller, slower production of misinformation. We just don't have laws in place like that. The last thing I'll say is that on my side of the Atlantic, Donald Rumsfeld famously talked about how there are known knowns and known unknowns and unknown unknowns. And we're really in a regime now of unknown unknowns where we just don't know all of the ways in which these systems are going to be used and misused. As I say, that's changing every day. Bad actors are digging in and it's hard to even know where the regulation should begin. But it is, I think, there's so much happening so far. fast that we need to take seriously regulation right now. However, today's debate might turn out, and let me say, I would like to lose. I would like to be proven wrong here. Whether I'm right or Jeremy's right, there's certainly going to be some bad, bad consequences, and we do need to think about how to regulate them. And I hope that people's love of technology won't blind us to that fact.

You're listening to our debate today, be it resolved, chat GPT will do more harm than good. Jeremy, maybe just worth spending a little bit more with you thinking about how this technology could evolve in the future. And particularly, you're on the side of this debate that these technology can be positive. So I think we all acknowledge that what we're seeing right now with chat GPT is the kind of first iterations of this technology. So where do you see it going? How do you see it evolving?

And then what potentially are the, I don't know, the personal, the individual, but maybe also the kind of collective goods that could emerge? from ever more robust large language learning models that may extend well beyond text to voice, to video, to visual images, and eventually maybe to the amalgamation of all those kind of sensory inputs, the visual, text, auditory, all combined together in these sophisticated machines. I think the scenario you lay out in which these systems increasingly understand more than just massive amounts of text. They also have some understanding of visual imagery, maybe of sound. That is already starting to happen, and it is clearly, you know, something that's going to

happen in the future. There's called multimodal models, and a lot of large models in the future will be multimodal. There are already people using them for very short video generation from text. So you might be able to just give a text description of what you want, and it will create a video, it's a pretty short video. You know, you're talking about no more than, you know, maybe 10, 20 seconds right now. But in the future, it's possible that you would even be able to create quite long-form video just from text descriptions. And I think, you know, that potentially opens up all kinds of creative venues. I'm sure Gary will tell you it also opens up, you know, yet more, you know, opportunities for creating, convincing misinformation. But it, but it definitely,

you know, might supercharge creativity. It might create, you know, enable people to create, wonderful films without needing a Hollywood studio to back them. So there might be a kind of flourishing of content in a way. And I think that could potentially be a benefit. I think you'll also see on the language side that these systems do improve, that there will be improvements on this issue of hallucination, that we will find ways to make sure that what the system generates is true or that there's a sort of reasonable assurance that it'll be true. that they will hallucinate less than the current systems tend to. Can I jump in on that one?

Yeah. Large language models by themselves are blind to the truth. Making them bigger doesn't make them more truthful. It makes them more plausible. I do think we will eventually develop new technologies. I have some thoughts myself about how to do so that are able to fact-check these systems. But it's a mistake to think that the current tech by itself is going to solve that even as the models get bigger.

What they do is they track the plausibility of words in context. They don't have an intrinsic way of representing some baseline set of facts and validating things against them. So, for example, a large language model might tell you that Elon Musk died in a car crash in 2018. That's something Galactica actually said, even though there's nothing in the database to support that and lots of things to go against it. They feel like they're smart, but they're not smart enough to validate things either against the internet or against their own facts. People are building other symbolic technology to fit in with that, but it's an uphill battle because you have a black box that is disregarding the truth and not telling you how it got to what it got. So this is a serious problem for a while.

It may not be a problem forever, but it's going to require serious advances in how we build our AI in order to solve it. I think, Gary, that you will find these things coupled, though, again, with other systems that do you work in a different way. And I know you yourself have been a big advocate for sort of neuro-symbolic hybrid systems. And I think, you know, we will see some of that. with large language models playing a part of that. And I think you will see sort of this, also some best people talk about sort of system one and system two to use, you know, Connman's taxonomy and that most of these deep learning systems are kind of kind of system one technology and that we will see other technologies used to provide a kind of ego oversight, you know, a kind of system two oversight of what these things generate. I mean, I think we absolutely need to go there. And that's a nice way to put it, is these are system one type systems, and we're going to need symbolic systems to give a system two. I think that's exactly where the field needs to go. And we need to be realistic that the field has been leaning against that for a number of years. And this requires a shift. And I'm hopeful about that shift in the long term. In the short term, because of all the investment has gone to the large language models, we're a little bit flat-footed right now. And there's going to be a few years of rough ride until we get that right. So Gary, just before we go to closing statements, what's your view? If we're, if we're

going to synthesize these different abilities and features of large language learning models together, voice, text, audio, we're going to maybe have to wait a little while for the, the ego to arrive to give this kind of shape and form. What do you see as the downside here? I mean, right now you've outlined some practical short-term problems or challenges, but what are you worried about longer term? You talked to the beginning of this debate. I was intrigued about this idea about the importance of truth to democracy and how maybe there are a series of unintended consequences, a kind of new tragedy of the commons that could unfold here around some of the very touchstones that we've built modern Western civilization society on. You can't really get

much bigger than that. I mean, I think we're leading towards a world that authoritarian's like, which is one where people trust nothing and therefore sort of need the strong man. I think that that's definitely an immediate risk. I think there's a little bit of a tragedy within the field that there's been so much investment in the large language models that have this problem of hallucination that was not really well recognized until recently, even though I've been talking about it for 20 years. that many other things like neurosymbolic AI have been excluded. And that's been to the benefit of a certain number of, you know, small number of people at major companies, but it's probably not been to the benefit of the development of AI or to

the benefit of society as a whole. And so, you know, I don't think right choices were made. Maybe we will start to shift the boat, but we're not in a good place right now. Jeremy, do you want to come in on this, this bigger point of kind of truth, democracy, technology. I would say that people might have made similar arguments, you know, about the printing press and other technologies about broadcast technologies when they came along, that these things, you know, would somehow obliterate the truth. They would make it very easy to create mass propaganda. And all that was true. They did make it easier to create mass propaganda. You know, you don't

longer needed a monk and a scribe to get your book out. You could just print it. People could say whatever they wanted. There was suddenly this explosion of ideas. And, some of those ideas were not true. But it actually expanded the potential of people to express themselves. And I think actually chat GPT may do something similar, that people who had no access to formulating things at scale, to creating companies, to getting their ideas sort of out into the actual marketplace, let alone the marketplace of ideas,

will have the ability to do so in part because they can do so with the help of these assistants that save them a lot of time, create a lot of efficiency, you don't need as many people to create a viable business out of them. And I think that could have a tremendous benefit to society. And I think also they are likely to create a healthy skepticism on the part of publics, at least publics and democracies, about information where it's coming from. And people will have to go back through and analyze things and look for sources.

And that's not an unhealthy thing either. So I think on a whole, over the long term, that these technologies will probably provide more benefits than they will harm. Okay, you're listening to our debate today, be it resolved chat GPT, those large language models that everyone's talking about, will do more harm than good. Let's go to closing statements. Gary, I'm going to ask you to go first, sum up this debate for us. What are the key points or ideas that you want to leave people with to, indeed, understand why? they should be thinking about large language learning models as something that could be quite harmful to society and the steps that we might want to take.

Well, I think I agree with Jeremy in a way around education that the way forward here is to understand the limits of the models and also to build new tools to fight off some of the dangerous things that they do. It really is a win if we wind up with a more critical society that can do better critical thinking and critical reasoning. But I don't think that we're super well positioned right now to do that. And so I am less optimistic that we can do that quickly. I think that the kinds of tools that we're building right now are better at generating misinformation than detecting it, that we're going to need new technologies for that,

that society needs to support that in order for us to cope. And the last thing I'll say is I hope that I lose this debate in the sense that I hope that in time we discover that there's not a serious. a problem with a large-scale production of misinformation as I'm anticipating. You know, I certainly hope that I lose, but I'm pretty concerned that the 2024 election will be the first major election in which this is a major, major part of the discussion and what goes on. Thank you, Gary Marcus.

Okay, Germancom, we're going to give you the last word in our debate today, be it resolved. Chat, GPT will do more harm than good. You've been arguing against the proposition. Wrap this debate up for us. So I think, you know, chat TGPT in the end and related technologies, these large language models, they just will become a sort of another tool that we are using for all kinds of things. And on a whole, because they are a general tool, they can be used for so many different tasks in so many different industries and professions that this will have tremendous benefit to society.

And again, to emphasize a lot of that has to do with efficiency and time savings. A lot of it has to do with empowering people to do more with smaller teams than they could before, which I think is, you know, it's good for startups, it's good for people who run their own businesses. And I think, you know, those people will see tremendous benefit from this technology. And we will see companies being built to greater size with fewer people, which, again, you know, is tremendous sort of economic benefit there potentially. And the hallucination problem that Gary talks about is real, but I think we will come up with strategies to mitigate this, in part by using this alongside other kinds of technology.

And again, I think if you look at how chat GPT and other large language models tend to be used by companies, they tend not to be used in isolation. So again, I think that the danger of some of the harms that the solutionation could cause that Gary has correctly identified, in practice, you know, will be mitigated by these other solutions. And so I think on the whole, and again, in the long term, these systems, will provide tremendous benefit and that the harms over time will diminish. Well, Jeremy, Gary, thank you for a really fascinating debate. Unpacked so many of the key issues and ideas that I hope we would in our conversation together

and just on behalf of the Monk Debates community. Thank you so much for your time, your insights, and for sharing it all with us today. What a great conversation. Thanks again. Thanks very much. And thanks, Jeremy, for joining. Yeah, thanks, Gary. Thanks, thanks, right. Well, that wraps up our debate today. I want to thank our participants, Gary Marcus and Jeremy Khan. What a terrific far-reaching conversation. If you have feedback or reflections on what you've just heard, please send us email to podcast at monkdebates.com. That's MUNK DebateswithanS.com. Also a friendly

reminder that you can catch right now in the same podcast feed a sample of our weekly current affairs podcast, Friday Focus, that comes out. out you got it every Friday with a half-hour dive into the big geopolitical stories and issues that are shaping the news. You can sign up for the full-length editions of Friday Focus on our website, triple-w monkdebates.com. Simply go to the top right navigation and look for Friday Focus. Thank you for lending your time and attention to our mission to bring back the art of public debate one conversation at a time. I'm your host and moderator, Rudyard Griffith. The Monk Debates are a project of the Aurea and Peter and Melanie Monk Charitable Foundations. Rudyard Griffiths and Ricky Gerwitz are the producers.

The Monk Debates podcast is mixed by Adam Karsh. Be sure to download and subscribe wherever you get your podcasts. And if you like us, feel free to give us a five-star rating. Thank you again for listening.