Jack Clark on AI's Uneven Impact | Conversations with Tyler

Well, thanks for having me, Tyler. Pleasure to be here.

I'd hazard a guess that it's going to be things that are the trades and the most artisanal parts of them. So you might think of trades as having things like electricians or plumbing, or also things like gardening. But I think within those, you get certain high-status, high-skill parts where people want to use a certain tradesman not just because of their skill but because of their notoriety and sometimes an aesthetic quality. I think that my take might be gardening actually. And they won't use AGI to help design the garden? The human front will never disappear. I think the human front will never disappear, and people will purchase certain things because of the taste of the person, even if that taste looks like certain types of modern art production where the artist actually backs onto thousands of people that work for them and they're more orchestrating it.

There are certain types of deskbound work that just require talking to other people and getting to alignment or agreement. You might think AI is great at doing that already—it's a wonderful therapist. But we don't send Claude to sell Claude yet; we send people to sell Claude, even though Claude could probably generate the text to do the sales motion. People want to do commerce with other people. So I think that there'll be certain relationships which get mediated by people, and people will have a strong preference probably for deals that they make on behalf of their larger pools of capital, where the deals are done by human proxies for large automated organizations or pools of capital.

I think a few years ago it was encountering quite strong obstacles in the law itself, because lawyers tended to like being able to charge very high prices and had a dislike of things that could bid them down. But my less flippant answer is probably big chunks of healthcare. It's bound up in certain things around how we handle personal data, all of the standards around that. All of those standards are probably going to need to be changed in some form to make them amenable to being used by AI, and we've had a really hard time updating or changing data standards in general.

It will change in the form of gray market expertise, but not official expertise. I had a baby recently, and so whenever my baby bonks their head while I'm dialing the advice nurse, I talk to Claude just to reassure myself that the baby isn't in trouble. I don't think we actually fully permit healthcare uses via our own terms of service. We don't recommend it because we're worried about all of the liability issues this contains. But I know through my revealed preference that I'm always going to want to use that. But I can't take that Claude assessment and give it to Kaiser Permanente. I actually have to talk through a human to get everything else to happen on the back end to work out if they need to prescribe my child something. So the number one job will be surreptitiously transmitting the generation of information that comes from AIs. In essence, some of it may be about laundering the information that comes from AIs into human systems that are not predisposed to that information going in directly.

I think that would be my forecast? They still use software from the 60s or maybe even the 50s sometimes. But I actually suspect that that could be a place where we see really rapid changes, actually, for a couple of reasons. One, we know that AI has relevance to national security—it develops certain types of capabilities.

Yeah, that'll happen quickly. But the rest of government—the Department of Education, HHS, the non-scary shot part of government—yeah, I would wager that it will become surprisingly easy to get it into really hard parts of government. And then there will be a question of political will. But all around the world, and I'm sure you've experienced this, governments desperately want growth and they desperately want efficiency. I mean, we see that here today. And they say that, oh, but I think if you look at also things like voter polling and other things, people want to see more changes out of government than they're currently getting. And I think sometimes constituent preferences do ultimately change what their elected officials do. I would just take the other side of this: that government may move slightly faster than you think—it may be very large established companies that end up having some of the greatest resistance to this in certain areas.

I think that it will only work in a scenario where the system has got so powerful that you can bring the AI system itself in to help you think through this. And then there will be a question of political will, which is where the system might all break down.

I think there is a high chance for a political movement to arrive which tries to freeze a load of human jobs in bureaucratic amber as a way to deal with the political issues posed by this incredibly powerful technology and how fast the changes are going to come. I don't think that we'll do this in a reasoned way; I think it'll be driven by the chaotic winds of political forces. And it feels like the sort of outcome that happens if we, as the companies building this stuff and our customers, don't generate enough examples of what good transitions look like. The fewer bits of evidence you have there, and the more evidence you have of larger economic changes, probably the higher chance there'll be a desire to step in and protect workers in different domains. That comes from an impulse based around wanting to help people, but it might ultimately not be the most helpful thing over the course of decades to do that.

I believe that all people have a desire for meaning that is salient to what they do on a day-to-day basis. And my worry with what you describe is it might not feel like it has sufficient meaning. I think that there is some giant class of activity that we want to continue happening in the world that people do and from which they draw meaning. But I don't know that the best way to get there is to take some class of work and say, 'This is work that we're protecting and from which meaning will spring,' because I'm not confident that you'll pick a load of jobs which naturally create their own meaning in that sense.

There's some of that, but even the Nobel laureates, they're rivalrous with each other. They can be very, very petty. But that's just human nature, right? If the Nobel laureates aren't happy, there's no post-AI world that's going to do much better than how we're doing for the Nobel laureates today.

I maybe my pushback is that I think that all of this could happen sufficiently quickly that we might have the opportunity to just play different higher-status games that are afforded to us via AI and the productive capacity it unlocks.

Yes, but I think that there are going to be analogies which look more like creative, fun exercises in getting AIs to build things or make things, or almost carry out competitions and games where people can play them with one another. And I think there'll be entirely new forms of entertainment that have some amount of meaning and perhaps an economic engine wired into it that people can participate in. I believe we're not that far from the age of what I call the AI teddy bears. You know what I mean when I say that?

I've had this thought since I have a person that is my child that's almost two. Sure. And I am annoyed I can't buy the teddy bear yet.

I don't know. I don't think I'm an outlier. I think that once your lovable child starts to speak and display endless curiosity and a need to be kind of satiated, you first think, 'How can I get them hanging out with other human children as quickly as possible?' So we're on the preschool list, all of that stuff. I also think—I've had this thought: I wish you could talk to your bunny occasionally so that the bunny would provide you some entertainment while I'm putting the dishes away or making you dinner or something. Often you just need another person to be there to help you wrangle the child and keep them interested.

I think that's the part where you have them spend more time with their friends, but you keep the bunny in their life because the bunny's just going to get smarter and be more around them as they grow up. If you take it away, they'll probably do something really strange with smart AI friends in the future. No, I don't think I'm an outlier here. I think most parents would, if they could acquire a well-meaning friend that could provide occasional entertainment to their child when their child is being very trying, they would probably do it.

I agree. Yeah, yeah. But the same as with screens, a lot of children they keep on wanting it. It's hard to tell the child you can't have it now. It's like in the old days, it was watching television. 'Oh mom, can I watch Star Trek again? Can I watch TV seven hours a day?' Well, that's not good, and that's hard to ration.

I mean, we do this today with TV where if you're traveling with us on a plane, or if you're sick, you get to watch TV, but otherwise you don't, because from various perspectives it seems like it's not the most helpful thing. You'll probably need to find a way to gate this. And it could be when mom and dad are doing chores to help you, you get the thing. When they're not doing chores, the thing goes away.

I think it comes down to frequency, which I mentioned, and also what you are and aren't allowed to know. I find surveillance puzzling. We have one of these smart webcams that we got when the kid was early so that you just see that they're asleep and you know if they're waking up, you work out if they're okay or not. And the main thing that that meant is that when the baby cries at night, sometimes we go in less than if we didn't have the camera, because if we didn't have the camera we'd have to go and check. And as a consequence, my baby tends to sleep through the night a lot more because they don't get occasionally interrupted. Sometimes they wake and cry for a minute and just go back to sleep. So I think sometimes this stuff allows you to actually interfere less in a person's life if you know certain things about it.

If the bunny says your kid's saying strange stuff? But they all say strange stuff. I might have been chattering on about the New York Mets at age seven and it would have been perfectly harmless, but it may not have sounded that way. You're going to need to create spaces for unmonitored creativity in people. Actually, the same as how we have an approach to AI research today. Now AI systems can output their chains of thought, which is the reasoning they use to come up with answers. And we've had this question at Anthropic of how much should we monitor the chains of thought. Because if you actually monitor them, you might create an anti-goal where the system ends up wanting to have chains of thought that are safe to be monitored and which don't cause demerits, which might actually break how it thinks in a bunch of ways. I think this analog applies here, where you're going to need to choose how much you actually decide to know about people, or you risk creating incentives that change their behavior such that they have a negative effect.

Although if a person asks you to do that on the first date, it's fine. And if they ask on the second date, you probably shouldn't be dating that person.

I met my wife on OkCupid, where you fill out a survey—of course Match.com for me, where I met my wife—and then I met her and our non-AI but automated system had said you guys might get along, so we met each other that way. So I don't know that this is so different to previous things we've used.

I just think that this is an area where we're going to figure out the new norms of this technology and what seems appropriate and not appropriate. And some of that is just going to be solved through the logic of business and other things through usage.

I think that this is one of the toughest questions in front of us. As you know, I'm a former journalist, and I grew up during the period when online ad market changes were altering the business model of journalism, because they were switching from value that you derive through quality through subscriptions or people buying your stuff, to value that you derive from just large-scale attention, because that became a lot of the model for funding this stuff. And as a consequence, you saw the need to cross-subsidize journalists with journalists that got loads of eyeballs, and journalists like me who would write about databases got less eyeballs, so we were cross-subsidized by the people that wrote about Justin Bieber or what have you. I think for media, it's going to be really challenging to think through how the economics of this work. I think that it'll change in a couple of ways. One, you might move to some of these larger scale publishing house style models for certain types of fictional universes which have subsidy and cross-subsidy within them. You see where does the cross-subsidy come from? Like Bloomberg, we know how that works, right? But if AGI is truly general and it's based on what's out there already, it should be able to do better than media on all dimensions.

Yeah, or Warhammer 40k, to show my nerd credentials. But both—they'll publish fictional news, right?

I think some real news comes from analysis of publicly available facts and it being composed together in a way that shows you insights that didn't exist. Semi-analysis on Substack is a good example of this, a lot of public stuff leading to interesting conclusions. But news in the moment that has loads of context was subsidized by previous business models which mostly no longer work, and I genuinely don't know what happens to it.

I would expect that there will be, for quite some time—many, many years—areas where there are slightly different specialisms on all of these things. There'll be concentric circles that have got massively large in terms of the space that they cover, and they'll have edges which are slightly differentiated. But in those edges will be where the frontier of certain types of human value stacks on top of AI value to lead to people choosing one or the other because one is more poetic, for instance—a bad example, due to the economics of poetry not being particularly great—but coding might be a better example. Or there could be cases for things like certain types of scientific experimentation where taste might come to matter a lot for composing experiments. I expect we enter a world though where you have a single-digit number of these very large-scale models which are servicing a much larger set of wrappers on top of them that change the form factor by which they get integrated into your life, and which have a load of helper functions or knowledge probably built by those AIs in partnership with some other specialized AI system or expert knowledge from that domain.

I think right now it's like we're car companies and we are mostly selling cars to teenagers who say 'How fast is it?' and 'Can I get it in red?' And we're beginning to find customers that are businesses that operate fleets of cars and say, 'What seat belts does it have? What is your accident rate? What are all of these other properties that the teenager has never asked us about, but these businesses do?' Because it ties to their business model or to liability or other things. I think therefore there are technologies to be built that look like the safety equivalent that we have in cars or other things, which will change the logic of competition. So we're in competition right now, and that competition will change as we unlock different markets that have different properties that need to be present in the AI system. And those properties will often ladder up to certain types of safety technology. So the price of insurance is carrying the mechanism, so to speak. I believe that this will be one of the ways that it helps.

I think that I have thought less about liability and more about information that should be made available by the companies. But I think these have an interplay where today there isn't really a common labeling or disclosure standard about what is in these AI systems or the industrial practices you've used when making them. And I think that this is ultimately an area where you can do interventions to get some common level of transparency, and that'll interplay with both liability and also things like negligence, which will change the behavior of these companies over time. So there's some level of common information we can start providing about these systems which will also change corporate behavior.

Well, one of the things that you do is you create information which becomes truly common. I mean, the economic index work we've done is from one single firm, but ultimately you might want to generalize that to all firms, but then link it to large-scale data gathering that governments might do. And I think if that became the case, you would have a higher chance of tying it to actual policy responses that would be larger in scope. I also think on climate change, just by measuring things like parts per million, that has helped catalyze large-scale amounts of capital to get remobilized around the economy in different areas to deal with this perceived negative impact.

90% agreement. I think there is a chance of something that looks like a non-proliferation agreement between states, including the US and China, in the limit, and the US and China enforce it and we become vaguely allied with them in this one crusade. I think it might be deciding that certain capabilities of AI systems might be so potentially destabilizing that you don't want them broadly available, while recognizing that each government may be privately developing them. But you might have some common standard for what's available to everyone where you've decided this would lead to all kinds of chaos in our countries as well as in other countries as well.

My assumption is a lot of this gets enforced by checking and then throwing policy punches at each other on things like tariffs or exports or other issues. I don't know that there are global governance bodies that will be formed. Some people feel optimistic about IAEA-style models for this, but I think things that require mutual inspection regimes end up being very, very hard to do under certain rivalrous dynamics.

So say 10 years out, what's your best estimate of the economic growth rate in the United States?

I think my base case on all of this is 3%, and my bull case is something like 5%. I think that you probably hear higher numbers from lots of others—20 and 30? I hear all the time. To me, it's absurd. And the reason that my numbers are more conservative is I think that we will enter into a world where there will be an incredibly fast-moving, high-growth part of the economy, but it is a relatively small part of the economy, and it may be growing its share over time, but it's growing from a small base. And then there are large parts of the economy, like healthcare or other things, which are naturally slow-moving and maybe slow in adoption of this. I think that the things that would make me wrong are if AI systems could meaningfully unlock productive capacity in the physical world at a really surprisingly high compounding growth rate—automating and building factories and things like this. But even then, I'm skeptical because every time the AI community has tried to cross the chasm from the digital world to the real world, they've run into 10,000 problems that they thought were paper cuts but in aggregate add up to you losing all the blood in your body. And I think we've seen this with self-driving cars: very promising growth rate and then an incredibly grinding slow pace at getting it to scale. I just read a paper two days ago about trying to train human-like hands on industrial robots using reinforcement learning. Doesn't work. The best they had was a 60% success rate. So if I have my baby and I give her a robot butler that has a 60% accuracy rate at holding things, including the baby, I'm not buying the butler. My wife is incredibly unhappy that I bought it and makes me send it back. So I think as a community we tend to underestimate that, and I may be proved to be an unrealistic pessimist here. I think that's what many of my colleagues would say, but I think we overestimate the ease with which we get into the physical world. As I said in print, my best estimate is we get half a percentage point of growth a year. 5% would be my upper bound.

I think one is something that you touched on earlier: we could get the politics of this really wrong, and the technology could just get put in a relatively small box that does some economic good somewhere but at a very small constrained amount. That's the nuclear power failure mode story, where we do some large-scale regulatory scheme, make it hard to do this stuff, it ceases to have much of an effect—maybe it has an effect elsewhere. The other case I'd give would be, it's very hard for me to give a 0% case because we know today it's incredibly useful for coding. And if you just stopped all of it today, all further progress, I think the coding use case alone is going to be incredibly useful because it grows the ability to digitize parts of the economy, which we know drives faster loops in different parts of the economy and generates value. So I think 0% is basically sub-1% chance of happening, and if it did it would be a lite thing or something that looked like war with Taiwan leading to everything else looking different as well.

I think that dense aggregations of humans have significant amounts of value, and I would expect that a lot of the effects of AI are going to be, for a while, massively increasing the superstar effect in different industries. So I don't know if it's all cities, but I think any city which has something like a specialism—high-frequency trading in Chicago, or certain types of finance in New York—will continue to see some dividend from sets of professionals that gather together in dense quantities to swap ideas.