
Jack Clark (Anthropic cofounder) argues AI progress will reshape economies, science, labor markets, and policymaking; governments must govern via transparency, safety testing, and infrastructure readiness, especially in developing countries.
Jack Clark is cofounder of Anthropic, developer of Claude (an AI-based chatbot). He speaks at the IMF World Bank annual meetings in Bangkok and is interviewed by Gita Bhatt (IMF).
AI can accelerate scientific R&D and create “automated science laboratories.” Policymakers may underestimate fast diffusion and AI-driven inflation in supply chains (PCs, RAM, components). Frontier firms should face mandatory transparency on safety/security (prevent theft, prevent creating vulnerabilities, test safety). Societal risks are harder than national security risks, especially harmful, context-specific interactions with vulnerable users.
solar panels and medical cures as historical analogies; rapid coding diffusion around 2024; “Claude Corks” as an experiment for work placement; Netflix-style friction to prevent excessive AI offloading; social-media failures (e.g., links to anorexia) as a caution for future regulation.
Written by AI. May contain mistakes.
Jack Clark shares his perspective on past fears and technological optimism.
“When I was a teenager, I read a lot about the world and I was afraid, as I suppose lots of teenagers are.”
Discusses the potential uses of AI in governance and societal betterment.
“to ask him what that fundamentally different world he speaks of might look like.”
Explores the underestimated impacts of AI on economies and labor markets.
“I mean, tech companies in some sense are experiments in this radical form of optimism.”
Strategies for developing nations to leverage AI for growth and infrastructure.
“So what can developing countries realistically do about that?”
Discusses obligations and regulations for frontier AI companies regarding safety.
“where in my experience, having done some analysis, these supply chains cover the top 50 % economies of the world.”
Examines the distinct human attributes that remain valuable amidst AI evolution.
“and what bad practices entail, then you do targeted rules.”
Discusses the personal responsibility and societal implications of interacting with AI.
“The other is something here about taste and intuition.”
Explore how AI can be regulated and the societal risks involved.
“I think the question is how we get that to happen generally, because when I explain this to people, they're like, cool, strange hobby, not very common.”

Get the gist, explore an idea or ask about something said.
On IMF Podcasts, what are the key points of this episode?
Automatic transcript. May contain errors.
0:01Jack Clark: Taking AI's progress seriously means we should give ourselves license to imagine a fundamentally different world. AI is not only changing the way we work, it's transforming economies. Jack Clark is the co-founder of Anthropic, which developed Clode, the AI-based chatbot. Here he's addressing an audience of some of the world's most influential economists and policymakers at the IMF World Bank annual meetings in Bangkok. When I was a teenager, I read a lot about the world and I was afraid, as I suppose lots of teenagers are. I believed that the 21st century was going to be difficult, that population growth, climate change, resource extraction and other crises would add up to a Malfusian collapse of the human species.
0:51Jack Clark: And that after rising to invent computers and the internet, we might consume ourselves into the ground. And then over the subsequent years, I saw time and time again how technology allowed us to alter these odds. How solar panels came along and became much cheaper, much faster than people expected. How our growing understanding of the human body allowed us to invent cures for things that had killed many people, including some of my own family. How we could do things that I had never even imagined but find utterly amazing. I work on AI because it can help us do more of that. And I work on AI because whether it does that is about choices we make.
1:37Jack Clark: Choices people here in this room will make about how we direct it and govern it. My colleague Gita Bhatt met up with Jack Clark after his speech to ask him what that fundamentally different world he speaks of might look like. I mean, the best possible outcome is where there's like some blob of AI being used in a commercially directed way in markets and, you know, for whatever businesses want. But there is a large chunk which is basically being used to go and intentionally push forward different chunks of science with government support and backing and plans to translate that into gains. So my dream is that we have governments, ideally internationally, working on these big scientific challenges and a coherent plan for how, as we succeed, do we bring this to the world.
2:27I think that's one. The other part is that we have found a way to actually intentionally grow chunks of the labor force or labor market in areas where we think there's just going to be a strong preference for people, no matter what happens to the technology. I mean, we already, in most of the Western world, like chronically underpay teachers. And we get away with that because people want to be teachers. But this is like a choice that we're making in a very strange way, and we could instead choose to compensate some of these roles higher and actually enhance the number of them. This is a general problem.
3:00So you kind of intentionally skate to where the puck is going in terms of the technology. And I think finally you would like use this to understand, I think, some of the larger philosophical questions about what it means to be human in a very fundamental way. I think we have technologies that let us know more about ourselves and know more about our societies than anything that's ever come along before. And, you know, if I'm a politician, I can't interview every single one of my constituents. You can do that with AI. And AI systems could actually summarize for you the actual views of people. And maybe your politics can change such that you can act on those views, right?
3:38You can have better policymaking. You can have better policymaking that's actually more representative. Anyway, this is all my, you asked for my, like, true idealized vision. Of course, all of that requires a willingness to sort of believe that this is true. I mean, tech companies in some sense are experiments in this radical form of optimism. We believe this is true, so we've done a load of this. Like, we have a core interviewer system and we interview people that use our systems and we use it to change what we do. We have a chunk of compute that we subsidize or give away for free to scientists.
4:09So, you know, we're trying to model all of this, but we're only a single company. So, there's this question of how do you do that holistically? So you've argued that AI progress could accelerate quite sharply as AI itself becomes part of a research and development process. But what do you think economic policymakers are most likely to underestimate? The impact on productivity, labor markets, concentration, or something else? And what should we be watching for now? I think we are going to be slow to realize how quickly it could impact both prices in the form of hitting supply chains and raising their prices.
4:45I think if you look at the build out of the AI supply chain, it has bled through to changing the prices of consumer PCs, changing the prices of RAM, changing the prices of components that go into it. And I think policymakers will be surprised by, why are we seeing inflation in these parts of the world and it will be attributable to not just the AI build out, but the science and robotics build out that's now going to be stimulated by it. The second part is, I think the suspicion of people working in AI is that diffusion could end up being unusually rapid. We saw this with coding, where it seems like AI systems had been becoming better at coding for a few years.
5:25In about 2024, they suddenly started to diffuse very rapidly. And now it seems to have upended and changed software engineering in a bunch of ways that show up now in sectoral level slices of data. So they'll be surprised by that as well, I think. So if the frontier moves much faster than diffusion, do we risk a world where a handful of firms and countries capture most of the gains and rents from AI, while others become largely users of technologies developed elsewhere? So what can developing countries realistically do about that? This is always for risk, but it's for the gist of my talk here at the IMF was really about how we can reduce that risk by basically pre-positioning the infrastructure and the platforms that will be accelerated by this.
6:09So we have a larger set of people that are set up to benefit from it. what policymakers can do in, say, developing countries is closely look at any part of their AI supply chain, science supply chain, or robotics supply chain, and assume that this is about to become more valuable and a source of leverage. So don't let those companies be sold to other countries, like find ways to support them and sort of plan for their success, because I think that's been one of the stories so far. The other is look for opportunities to leapfrog. I mean, Around the developing world, we saw some nations leapfrog to 5G telco infrastructure, which has been massively beneficial to their economies.
6:49The same is going to be possible now with things like how AI can upend healthcare or it can upend other systems, or it's going to create this new breed of automated science laboratory, which we already see the early signs of. Build that now, and it lets you take advantage of the gains when they arrive, while existing institutions are going to be much slower to convert themselves. So if you were advising a minister of finance, a low-income country, with only$50 million to invest in AI readiness over the next five years, how would you spend it? Okay, I wouldn't spend the money on power plants or data centers.
7:24Those things are wildly expensive relative to that budget. I would set like$1 million to$2 million aside for just measurement and telemetry of labor markets so that I could see what was happening and be able to attribute that to AI or otherwise. I'd set 8 million aside for young, like early graduate, basically work placement schemes. Anthropic is running an experiment here called Claude Corks. I think that that is like a valuable experiment for many nations to run. 40 million, the rest, I would do for investment into either existing firms or ways of capturing the upside from the AI science or robotics supply chain, where in my experience, having done some analysis, these supply chains cover the top 50 % economies of the world.
8:15So top 50 half of the countries of the world have some relevance here. It may extend even deeper. So I think you can always find a thing. It just requires foresight. Right. So now if capabilities are advancing faster than government's ability to regulate them, What obligations should frontier AI companies subject themselves to? And which of those should be mandatory rather than voluntary? There should be mandatory transparency about the safety and security practices of the company. By this, I mean, how are you ensuring that your stuff cannot be stolen? And how are you ensuring that your stuff can't create security vulnerabilities elsewhere?
8:54You know, stop people being able to steal your stuff, because if it's ahead of anything else, that's dangerous. stop your stuff being able to steal other people's stuff. So make sure it can't break out of the system. Also, safety, you know, how are you testing your systems for safety and what are the properties that you're discovering? And the reason that should be mandatory is it creates a body of knowledge that the world can then use to actually design more specific regulations where ultimately we will be able to attribute things that happen in the world to the development practices of companies and we'll be able to realize what good practices seem to prevent and what bad practices entail, then you do targeted rules.
9:36What regulations can we really cooperate on? I think we can cooperate on this notion of transparency because it doesn't require specific expertise or knowledge in regulators. It is just a basic sort of almost market access requirement. And I think you can arrive pretty quickly at a global consensus of the basic things that everyone needs to share. And then ultimately, you can turn that into more toothsome regulations. Secondarily, different countries in the world have now built AI security institutes, ACs, that are often groups of technical people that can do testing and evaluation of AI systems.
10:19Many countries have these, not just the US and UK, and China has a federated version, but others. These can clearly allow us to coordinate on building different kinds of tests, and they may each bring their own expertise. So to avoid duplication of effort, you may want to encourage specialization among the ACs. And then you can have a kind of test once run anywhere where you agree, oh, this AC's got this competency, go to them for this form of testing and vice versa. So are you surprised by the power of recursive improvement? I mean, you know, Mythos is now already old technology. If machines increasingly perform both physical labor and cognitive labor, what becomes a scarce human contribution?
11:07Is it judgment, intuition, creativity? What is it and what does it mean for societies? I would say there are two genres here. There is relational, where people will have a strong preference to relate to one another as people. And, you know, even though machines may be able to do lots of it, there are components of some jobs, especially in education and healthcare, where AI may end up automating large chunks of it. But people will have, I believe it as a parent, like a huge preference for other people to be in those roles. If I could send my kids to the only robot school, and it has, for some reason, better educational outcomes than the school of humans, I think I still probably send them to the school of humans, because otherwise I worry that it's going to do something very inhuman to them.
11:52So I think we'll have that. That's one genre. The other is something here about taste and intuition. So let's say even kind of creativity becomes something that machines become good at. I think that people have a good sense of the cultural, ethical, and moral norms of other people, which is constantly changing. Culture is constantly evolving. And our own ability to exercise good judgment there is going to be like a comparative differentiator and advantage. So, you know, my vision of some of the future when I think about work and I also look inside Anthropic is teams giving large chunks of what used to be seen as the work, like conducting research or doing direct technology development to machines, and us exercising some kind of judgment around how much should we know about this?
12:45Let's build systems that give us knowledge over it. And at what point do we want to set up intervention points or checkpoints on these increasingly automated systems? I think that's something where people are always going to have strong views, and you always want people to be making those decisions. So at these meetings, we heard a wide range of views about AI from optimists like yourself, putting it mildly, but others who saw nothing but cognitive decline everywhere. Are you worried about that? And what is it, again, for society and ethics? We can individually take actions here, right? Like I write this weekly AI newsletter.
13:25None of it's written by AI. It's all written by me. and the practice of the newsletter is me reading research papers and seeing what I've understood from them, writing summaries, and then I use AI to say to the AI, hey, have I understood this paper correctly, et cetera. I do that partly because I want to provide an educational service to the world, but also because it is a way of every week exercising my ability to have attention, to be able to read long things without any intermediary, and to be able to think for myself outside sort of the tool domain of like, what does this mean? So that's what I do individually.
14:02I think the question is how we get that to happen generally, because when I explain this to people, they're like, cool, strange hobby, not very common. I think at a technical level, we can put friction into the product. So when you watch Netflix too long, it stops at some point and says, are you still watching? I think that we can do exactly the same things for AI systems, where if you talk too long, maybe it says, hey, maybe this is enough talking to me, you should talk to another person. Or we've built a technology where AI can do a structured interview with you. So you might imagine that if someone is talking and seems very confused about a topic, the AI might say, hey, let's do a structured interview and see how much you actually know about this, because that engages you, it engages your agency, and it stops you doing the offload.
14:48And then finally, at a policy level, we don't provide services to under-18s. We recognize that kids are smart and will find a way to use this. But I think it's because we probably expect that there's this critical development period where I think most best practices say, don't throw too much technology at children when they're young. But what are other near-term risks that you see? And what should policymakers do then? Yeah. There are societal risks and national security risks, and they basically need to get treated very differently. National security risks are really, do you have capabilities which could be weaponized by other nations or non-state actors, cyber weapons, bioweapons?
15:35For that, you need a pre-deployment testing regime, and you need to make sure that you don't bring things to market which allow weaponization. and that to some extent we know how to do. The industry does that. The second is harder. It is the societal thing where let's say I am having a conversation with a chatbot and I'm in a vulnerable place mentally. And then because of that conversation, I do something tragic. I harm myself or I make decisions which turn out to be bad as a consequence of a machine. Like I take out a giant credit card loan because I'm manic and I've talked to it. I think that's a lot more difficult
16:12Jack Clark: to actually deal with because it's not like you can build a standard set of tests for this capability. It unrolls in context, it unrolls during the course of the conversation, it's unique to the person. My sense is that we need to build technology to be able to see in aggregate what sorts of conversations are happening and then design intervention points there. But this is a harder scientific challenge. And we failed to do this with social media. Like we knew on social media, say, in the image sharing apps, that it had some relationship to anorexia among young women. How do you create regulations to prevent that is something we haven't solved there, but we're going to have to deal with that same problem multiplied in other areas here.
17:02Jack Clark: So I'm reminded in all of this by the great Carl Sagan, who said in The Pale Blue Dot that the humans who one day may reach the stars will be a species very much like us, but with more of our strengths and fewer of our weaknesses, more confident, far-seeing, capable, and prudent. That's what AI demands from us. It's what AI demands of me and companies that I represent. And where Sagan thought we had hundreds of thousands of years to achieve this, to become a better and wiser people, We instead have a handful of years, less than a decade easily. I believe we can do it, but it's only by treating AI with the seriousness it deserves that we will succeed.
17:52Jack Clark: Thank you very much.
17:59That was Anthropic co-founder Jack Clark speaking at the IMF World Bank annual meetings in Bangkok in October. The interview you heard was conducted by my colleague, Gita Bhatt. Look for the Café Economics piece based on her conversation with Jack Clark in the December issue of Finance and Development magazine. Go to imf.org slash FND. And look for other IMF podcasts on Apple Podcasts, YouTube or wherever you listen. You can also follow us on X. Our handle is at imf underscore podcast. I'm Bruce Edwards. Thanks for listening.
We often fear what we don't understand, and the growing number of AI skeptics reflects the age-old struggle between man and machine. As a teenager, Jack Clark also fretted about the future until he discovered how technology could change the odds. Clark is a cofounder of Anthropic, the tech giant behind the generative AI chatbot Claude. In this podcast, he says AI can teach us more about ourselves and society than anything we've had before.
Read the article in Finance & Development: IMF.org/fandd