A (4:30)

Do you worry about how we're handing over too much judgment to machines? And what does that mean for how we arrive at shared ideas and beliefs of what's true?

B (4:41)

Many of the problems that those of us that have studied the sort of fracturing of the shared information ecosystem, we can disagree on certain ideas and have opinions, but there's a shared set of truths that we can all agree upon which forms the basis of the arguments that we have. And we've seen how that's become fragmented and fractured over the last 10 years. And in some ways you can see that large language models could make that worse in certain respects, because here's something that a lot of people don't really understand about large language models. If you, Tony, ask a large language model who's the best football player on the planet, and I ask the large language model who's the best football player on the planet, there's a pretty good chance we'll get different answers to that question. If you ask it about Universal Basic Income and I ask it about Universal Basic Income, there's a very good chance, so there is a real chance here that subtly everyone is receiving fluent and coherent answers that are actually quite different about the same subject, each one tailored to us. Now, you could say that's been happening with social media, but this is happening, I think maybe in a more subtle way because it seems to come from a super intelligent machine that's brilliant at everything. So we're probably more likely to trust it, keeping hold of our own critical faculties. When we have a super intelligent machine that is better at us, that at a lot of things, things might be one of the biggest challenges that we all face.

A (6:11)

A lot to unpack there. And your new book is all about how to talk to AI and how not to. So tell us more about what people misunderstand about using chatbots and AI tools and how can the way we interact with them shape the kind of knowledge we walk away with.

B (6:30)

Some of us massively overestimate what they're capable of. Some people call this the halo effect. We see that they are capable of producing incredibly long, sophisticated answers about physics or generating amazing poetry that, you know, you or I never would. You know, they can pass the bar in under a second and they can pass nearly all the medical exams. And so clearly in certain things are amazing. And I think that can lead people to assume that they'll be able to do absolutely every task amazingly well. So if you, if you ask it for some advice about your career, it must know better than you do. If you ask it about your area of expertise, it must be right. That is not true. That is not how they work. So we overestimate them sometimes, but I think some of us underestimate how good they are. We get a bad answer from them or they hallucinate and make up a reference or make up a fact and we decide they're terrible. They can't be trusted to do anything at all. Chuck them in a bin. But they help me immensely in writing a book. They are incredibly creative. People sometimes assume that a machine, because maybe it's hardwired into us from years of sci fi movies, that these machines are very good at facts and data, but not very creative. I think it's almost the opposite. They can have a level of emotional sophistication and maybe manipulation that would stagger you if you know how to talk to them well. So, so much of this is just about trying to understand what they're good at and what they're not good at. And a lot of that just comes through practice and open mindedness.

A (8:09)

Super interesting as a writer to hear how AI has, has helped you in, in writing your book. At a time when trust in traditional experts is already fragile, how should we think about the role AI plays in shaping authority and expertise?

B (8:25)

What I think a large language model has done is to break the relationship between fluency and coherence and style and genuine understanding and knowledge and insight. Language is so intrinsic to being human. When someone speaks with you with great coherence and authority, they must know what they're talking about. When you receive an email that's very carefully written and beautifully put together, they understand the subject matter. That relationship has basically now been broken that when you hear or read anything that sounds very fluent and very stylish or very well constructed, it does not any longer signify understanding because a machine could have just generated it. And I think this is one of the most important reasons people get trapped by machine hallucinations, machines making up stuff, making up facts. You know, it could even, it could be one small reference in an academic paper that's made up by a machine. It could be an entire universe of confabulated nonsense. We have to break in our own minds the idea that because this sounds authoritative, it is authoritative. That is no longer true. Accuracy and style are now separated. This is vital for everyone to understand. Doesn't mean to say there's not huge potential. I think for too long lots of people are locked out of brilliant ideas, either reading them or expressing them. These machines are actually very, very good at turning complex language into formats that you can understand, that anyone can understand. So there's like amazing democratic potential here for these machines to make everyone able to understand the world better and express themselves better.

A (10:23)

Looking ahead, what actually helps people navigate all of this better?

B (10:27)

There's a couple of things to know about machines that are just always worth keeping in mind. They are statistical, next word prediction machines. They don't have any kind of ground truth or knowledge of the world like you do, no matter how fluent they seem. So understanding that is essential in terms of always being slightly skeptical about what you're getting back. They are liable to hallucinate. Sometimes that's fine. When I am using it to help me with creative thinking, I want it to hallucinate, I want it to come up with crazy stuff. But if I'm reading through a contract and I'm about to sign and it tells me something incorrect, that could be a big problem. So the more important the task, the more careful you've got to be. And always bear in mind, you know, these are run by large profit making companies. They want to keep people online, so they tend to be agreeable. They will often flatter you. It will tell you it's an amazing idea. Tony, you're a genius. We are very susceptible to that, I'm sorry to say, which is the reason the machines do it. Especially if a machine's telling you something you might want to hear and makes you happy. Be doubly skeptical of this, okay? Like it's a flatterer. It can be a sycophant, it can hallucinate. If you know that it can be very, very useful. That's about the machine. But here's the other side of things, Tony. It's a two way conversation. When I put a question into ChatGPT, there's a very good chance that I'm smuggling in some bias of my own. So you gotta look inside yourself as well and guess What? One easy way to do that is ask a machine, how can I use you better? Hey, you know that question I just asked you about Universal Basic income? Was that an objective question and it might tell you no, that wasn't, but I didn't bother mentioning it. But now you've asked. No, you smuggled a premise in there about why it was a good idea and I gave you the answer. But if you want an objective answer, here it is. Human bias is almost as much of a problem when talking to a machine as, as machine bias. When you put those two things together, it can be really damaging. If you understand both of those things and you approach it with some skepticism and practice open minded, they can be really, really useful.

A (12:46)

Jamie Bartlett there on knowledge in the age of AI and how not to be fooled by it. Here at the World Bank Group, knowledge isn't abstract. Besides finance, it's the bank's stock in trade. And it deploys knowledge at scale, taking practical insights from one place to tackle global challenges like creating jobs. Pascal Donohue is the new Managing Director and Chief Knowledge Officer of the World Bank Group. Before this, he was in politics, where, as you know, evidence, narrative and public trust often collide. So I asked Pascal what drew him to this new role and why knowledge may matter now more than ever.

C (13:31)

Thanks very much, Tony, and great to be on your show. I mean, what drew me to the role is the opportunity to continue in public service. When I was a politician, I always viewed myself as a public servant. I'll be an elected one. And one of the big themes in my life in Irish politics was the importance of cooperation between countries to deal with the challenges that we were collectively confronting. And that experience, of course, made me so aware of the challenges and opportunities the world is confronting and therefore the importance of the World Bank Group. And it's such a great privilege now to be working for this great institution and to be working with so many exceptional colleagues in doing our work and work that has never been more needed. But of course, we're so aware of the challenges in doing that work and you've just touched on one of them there, which is the political and social environment within which our work happens. As an elected politician, I was so aware of it. And one of the big things I took from us was the importance of results, the importance of being able to show the policies that work and they make a difference. And using the theme of policies being effective, often working and being evidence backed, I found to be a very powerful antidote to some of the challenges that you've referred to. And of course that's really relevant to the work that I'm doing here now as knowledge officer for the World Bank Group. Because what we're trying to do is make the case for those policies that are very effective, that deliver better results on how we can do them better.

A (15:12)

So the World Bank Group talks about itself as a knowledge bank as well as a financing organization. Why is knowledge so central to the World Bank's work today, especially when it comes to sharing ideas that can actually scale and helping to create jobs?

C (15:27)

And so as a governor of the World bank and as somebody who looked on at the World Bank, I actually very much thought of it in terms of that concept. In other words, that there's kind of two streams of the World bank, if you wish, flowing together. What happens with finance and what happens with knowledge. And what I've really been struck by in the few weeks I've been here is the degree to which they're not parallel streams, they're actually streams that a lot of the time are already integrated and flowing together. That for colleagues who are directly involved in financial investments, whether it be loans, whether it be grants, they were always doing that in an environment that was very knowledge rich. And likewise for my colleagues who would have been working mainly on policy areas and mainly working on a more kind of knowledge track, they were always deeply aware of the importance of financial interventions for making sure that their knowledge go further. So I very much look at how we can integrate that work even more so together. And the reason why the Knowledge bank has this particular importance is because the needs that the world has are only going in one direction. They're growing. The opportunities are so great as well. But we're also aware of the many challenges that are there now with the availability of resources and the availability of money. And that's why knowledge has to go that bit further in a world in which resources are more constrained, but the needs and opportunities are growing.

A (17:00)

Given your formal life as a politician, I think you're well aware that we're living in a time where facts feel more contested than ever before and misinformation seems to move faster than evidence. So when we talk about development, how much of a risk is that for development? Especially in low and middle income countries where trust in public institutions is so important.

C (17:23)

So while we do face many big changes at the moment, it's not the first time we face them. When big changes have happened in the past in relation to communication, it has always heightened the vigor of public debate, whether that be the printing press Whether it be the availability of other forms of communication in the past, it led to more contestability and a higher level of vigour within public debate. I think what's different within my lifetime is you're seeing what I believed to be and still believe to be very objective elements of knowledge. You see more challenge and more debate around that. And I think in terms of what that means from a development policy point of view, it's firstly, we just have to acknowledge that this is happening and by acknowledging, then work harder to make sure that the claims and arguments that we are making with our policies are backed up by two things, firstly evidence and that that evidence is more accessible. And secondly, to go back to your point earlier on, that it's backed up by funding, it's backed up by resources. So we are saying these are not only things that we are making the case for our clients, for our countries, for our governments doing. We're also saying by the use of our balance sheet, we can make it easier for that work to happen. So it's evidence and balance sheet, evidence and balance sheets.

A (18:52)

Very well put. Now I'm curious to learn about your thoughts on the role of AI. In some places cases, AI can make a lot of these challenges much worse. But there's also an upside. AI can help people cut through the evidence and to make more sense of the evidence. But when you look at AI, do you see it as more of a risk, as an opportunity, or maybe a double edged sword?

C (19:15)

So I've always been a techno optimist and at times you kind of have to strain to maintain that fate, given some of the things that we know can happen. But, but I still am. And that techno optimism in particular comes from the role that I believe AI can play in healthcare and education in particular. I'm really aware of the challenges that we are going to face with regard to the future of work and the future of jobs. And the World Bank Group is doing some very important work in this area and we'll be bringing forward some research and some analysis. But when I look at health care and when I look at education from a healthcare point of view, I believe the availability of medicines, the availability of life science solutions that can deal with some of the great health conditions and challenges of our time, I'm really optimistic that that's going to change. And the other great challenge we face in health policy is the availability of nurses, the availability of doctors, the availability of consultants. And I think AI is really going to help with that, really going to help us come up with new ways of diagnosis. And then from an educational point of view, I again think that AI will be able to deliver great educational opportunities in the time ahead. Again, where schools where teachers are constrained, but they're not always available in the way we would want. So on balance, I believe the opportunities are great. But we have to be and are aware of the challenges and risks that it also does pose.

A (20:51)

Of course, when the World bank group talks about AI, we also talk about the responsible use of AI. So I'd love to hear from you what that means in practice and how do we ensure that? Or how is the World bank group using AI and knowledge systems to support better decisions while still making sure that it doesn't push out local expertise and institutions that they're strengthened rather than being sidelined?

C (21:14)

We have to see AI from a knowledge point of view as a way in which we can curate and organize knowledge. And if we think of it in that way, and it's the way I think about it most internally, it is then never playing a role in displacing where the knowledge comes from and from the humans who create it. And one of the things that I want us to see and do better is the different tools that we have, like knowledge, 360, like data, 360 that are digital and fundamentally AI tools, they're not the origin of the knowledge. The origin of the knowledge is the human, the human in the loop. But what they're seeking to do is organize that knowledge in a way that's dramatically more accessible than it otherwise would be. So that's really the way I would see it at the moment.

A (22:05)

Thank you again, Pascal.

C (22:06)

Thank you for having me on and

B (22:07)

have a good day.

A (22:11)

So there you are. Information is everywhere, but knowledge tested, shared and applied is what actually changes lives. Thanks again for listening to the development podcast. I've been your host, Tony Carasani, and I hope that you've enjoyed this episode as much as I have.