OpenAI’s Chief Economist on AI, jobs, and the economy | Ronnie Chatterji

This transcript is generated with the help of AI and is lightly edited for clarity.

//

RONNIE:
During COVID, when I worked in government, you had unemployment triple between, I think, the beginning of 2020 to maybe March or April. Pretty dramatic. But that’s what it took for unemployment to 3x. To make the same claim about AI in the next 12 to 18 months, I think you’re looking at a different economic model than I would be.

ARIA:
Are there other sort of corners of the economy that you think are going to adopt AI the quickest?

RONNIE:
Sure. I think that legal and people, or HR, are two orgs that people should not sleep on.

Commerce has become a major player, maybe the major player, on economic policy.

And I think if we tick down that list, we’re going to find that we’re better prepared for whatever comes.

//

REID:
In 1987, the economist Robert Solow made an observation that became famous. You can see the computer age everywhere except in the productivity statistics. Offices were full of PCs, and national numbers barely moved. The productivity boom finally arrived in the late 1990s, almost a decade later. And once it did, few doubted the computer had played a real part.

ARIA:
Today we have one of the economists best placed to tell us whether AI is about to repeat that pattern: Ronnie Chatterji, OpenAI’s first chief economist. From inside OpenAI, his team can see more of how people actually work with AI than almost any statistical agency in the world.

REID:
Ronnie’s colleagues at OpenAI built a test of AI against real work, written and judged by experienced professionals across 44 occupations. And Ronnie expects the bigger payoff to reach the productivity numbers within the next year or two. That’s a much shorter lag than the computer age gave us.

ARIA:
Our conversation with Ronnie unpacks how AI might shape the economy. It’s about when the change becomes visible, who it reaches first, what kind of economy we adjust into, and whether the companies building the technology can choose where it lands.

REID:
Before we get into it, if you’re watching on YouTube, give this one a like or subscribe to our channel. It’s the best way to help more people find these conversations. Now back to the episode.

ARIA:
Ronnie, thank you so much for being here today.

RONNIE:
It’s great to be here. Thanks for having me.

ARIA:
Absolutely. So you are OpenAI’s first chief economist. What did you think the job was when you took it? And then what has it actually turned out to be?

RONNIE:
Well, I remember my wife asking me, you know, what is this job you’re going to? It wasn’t something I was looking for. I mean, a lot of times in life I’ve discovered there are things that are great fits for you, but you never even knew you were looking for them. And when I came back from my government service, I had played a good role in terms of running the CHIPS and Science Act implementation. And so originally, when OpenAI reached out, it was about the chips. And we had a really interesting discussion about semiconductors and their role in artificial intelligence, but I really didn’t know if it was going to go anywhere. And we kept talking and talking and talking. And this is maybe typical of OpenAI.

And the conversation partners changed a couple times, but in the end, we sort of circled around to the idea that they need a chief economist. And I said, you know what? I don’t want to do this job, but I will write you a job description. You find the right person, they should have these credentials. And my life had been going back and forth between academia and public policy, and I wasn’t necessarily sure that I wanted to work inside an organization that way. Of course, I’d love to be a consultant. I’d love to have a say—all the great ways that kind of academic economists stay involved.

And when they came back to me and they said, look, we really think you’re the right person for the job, I said, okay, send me the job description. And it was the same job description that I had written. So that’s when I knew that I would have a chance to really shape this in a direction that I enjoyed. And when I thought about the promise of AI and the impact on the economy, my wife and I decided that this was an opportunity we couldn’t miss. So that was two years ago. So it was a great choice, and it’s been a wild ride.

ARIA:
And are you doing the same stuff from that job description today that you wrote up two years ago?

RONNIE:
Aria, I think it’s changed, like, six times since then. I mean, I had some things right. You know, I knew that I really wanted to do research that was beneficial to humanity, that gave people information about AI. You know, when you do a job like this, one of the things you have to think carefully about is, you know, you’re gonna have a lot of attention on you. You’re gonna have people paying attention to what you put out. And you have a duty, I think, in a role like this to make sure people have the right information on the impact of AI on the economy—something I know we’ll talk more about.

The second thing: I knew I needed a team, and I wanted a team full of people with PhDs and data science backgrounds, but also people who could manage, you know, inside an organization. I’d worked in government. I knew that getting things done cross-functionally could be really difficult. I wanted to hire MBAs, people who had sort of put things through the process. And so we have a mix of people who you might call program managers, data scientists, and economists. And that makes it—something I saw really early on that we needed.

ARIA:
Fantastic.

REID:
So your team publishes original economic research. So let’s kind of start the journey there. Looking ahead the next 18 months, what changes are you actually anticipating in the economy, and what’s kind of misleading noise in the general discourse?

RONNIE:
I think a couple things that I think will happen and are already happening in the next 12 to 18 months. Number one, we are in the early innings of an agentic revolution. It might surprise people to say early if you’re in San Francisco, but for the rest of the world, we’re just at the very, very beginning of this. If you look at firms that are using AI, there’s still a relatively small portion who have even turned on something like connectors—the ability to connect AI to existing software applications. That’s a big unlock. I mean, for all of us—we were talking about this in the green room—when you use AI and it’s not connected to your computer, if it’s just basically an ask-and-answer machine, you’re not getting the full benefits of that.

So I think that the agentic revolution is going to take over the rest of enterprise over the next 12 to 18 months. More firms will adopt. You’ll see it around the world in a big way. Second thing is, the job market hasn’t changed as much as people would have expected given AI capabilities have ramped so much. What has changed, though, is the work underneath the hood. When we look at what people are actually doing at work, we see a large percentage of work is now outside of your job description. So let’s suppose you’re a finance professional. Now you’re doing marketing, you’re doing people org work. Let’s suppose you’re a person in communications. Now you can run the numbers as well.

People are stepping outside their job role and doing things they weren’t doing before. You’re going to see that trend accelerate as more people have agentic power inside the enterprise. These are two things that are going to happen. The last thing I’ll say: I think AI will start to show up in the productivity statistics in a more significant way. It’s already showing up in GDP, as we know, through the capex channel—massive investment in hardware. But productivity is really what economists are looking at. I think that’s something you’ll see in the next 12 to 18 months.

REID:
And what do you think—and the other part of it is the noise question, because there are all these things that are said. Maybe it’s an overexpectation of how much AI is hitting already. Maybe it’s—what are the things you think, in the next 18 months, that are more misleading?

RONNIE:
I think that when people think about growth numbers, especially if they’re outside economics, they will cite really big numbers. Now, I worked in government during a couple really eventful moments, and, you know, growth rates of 10% are very, very high. And I think when people think about the power of AI, it’s easy to say, well, tell me what growth has been historically and multiply that by 5. Even 1 percentage point more growth in the US economy would be transformational. 2 percentage points, you would feel. And so the idea that we would get 10% growth in the next couple of years, that’s probably something that we’re not calibrated correctly on. Same thing on the job market.

You know, if the unemployment rate was to rise even a half a percent to a percent, there’d be a serious discussion about that. It wouldn’t be something that would just happen without us noticing. The idea that unemployment could triple in a short amount of time, like what you’re talking about, I think, is relatively unlikely. During COVID, when I worked in government, you had unemployment triple between, I think, the beginning of 2020 to maybe March or April. Pretty dramatic. But that’s what it took for unemployment to 3x. To make the same claim about AI in the next 12 to 18 months, I think you’re looking at a different economic model than I would be.

ARIA:
So we’re going to get back to some of the jobs numbers and sort of how we see it in the productivity growth you’re talking about. You looked at AI use in the enterprise. It’s increasing. No surprise. There are some power users, and there are some people in companies who aren’t using it at all. We’re sort of used to seeing that spectrum. San Francisco’s probably leading; engineering’s leading. Are there other sort of corners of the economy that you think are going to adopt AI the quickest?

RONNIE:
Sure. I think that legal and people, or HR, are two orgs that people should not sleep on. What I saw inside OpenAI is—you know, and our lawyers maybe are a selected sample—their ability to use AI to improve their workflows has been very interesting to me. And they’ve accelerated as fast as any other group I’ve seen. If you look at the data we had, November 2025, this is when everyone’s group chats, like, in San Francisco were heating up. Everyone was sharing all the amazing things. We remember this moment, right? This is not that long ago, but it really wasn’t until April or May of 2026—ancient history now—where the rest of the OpenAI org kind of got as agentified as our research and tech org.

So if you look at late 2025, pretty much all the research and engineering people were using agentic AI for 100% of their work. For the rest of us—and I’m sad to say as an economist, I’m part of the nontechnical workforce. All this education, all these degrees, all these numbers, I’m a nontechnical person now. But for people like me and others, we kind of got sort of fully agentified by April or May of 2026. That is about a four- to five-month lag between the nontechnical and the technical part of the org. I mean, Aria, that’s a pretty strong benchmark. I don’t think every organization is like OpenAI, but I would think the legal and the people functions, where there’s a lot of structured data and process, are places I would watch really closely.

ARIA:
Yeah, that makes sense. Structured data.

REID:
So one of the things that’s very Silicon Valley—here we are, one of the hearts of Silicon Valley, as we record this podcast—is starting a company. What does AI do to the cost of starting a company? Is it cheaper experimentation that creates more companies and jobs, or cheaper failures? And then, obviously, not just the creation of companies, but also scaling journeys?

RONNIE:
It’s a great question. In fact, I started my academic journey here in Berkeley studying entrepreneurship. And so it was an earlier era of entrepreneurial experimentation, but I decided to kind of devote my dissertation to studying this new thing, entrepreneurship, which wasn’t really a mainstream topic, but it seemed very relevant when I looked outside my window. I come back here now, years later, and AI is yet another accelerant of the entrepreneurial process and new company formation. I love this point you make, Reid, about more sort of rapid failures and why that can be good for the economy. When I first studied entrepreneurship, I always thought that the end goal was not a zero, but a one—the idea that you get a company right. But actually, you talk to people who’ve gone through their journey.

You guys have both done a lot of this. Sometimes the failures are where you learn the most. Sometimes the failures are the precursor to another great company you’d found. And sometimes, for many of us, we learn from that entrepreneurial experience that I’m better off working for somebody else, right? Most of us are more joiners and co-founders than we are founders. And so the entrepreneurial process—the quicker you can get to a resolution, whether it’s failure, and it’ll mostly be failure, the better. So I actually think one of the great things about AI and experimenting in entrepreneurship is it helps you go through the process faster, and that’ll be more failures, but also more beneficial successes too.

ARIA:
I also think there are so many—I mean, honestly, I think of my mom. She is a piano teacher; she’s a sole employee. She’s, like, technically an entrepreneur, and, like, AI can help her on her entrepreneurial journey in a way that she could not be helped five, 10 years ago.

RONNIE:
And she can focus on things that I’m sure she probably enjoys more, like working with students and teaching them, rather than acquiring new customers or figuring out billing or, you know, dealing with the parents, you know.

ARIA:
Some of us parents, yes. Every teacher knows that—don’t deal with the parents.

RONNIE:
But—but I do think you’ll see solopreneurship also enabled by AI. What we’re seeing in a lot of our data is people who are in parts of the country where they might have less access to a co-founder—hard to hire a chief marketing officer in some parts of the country. If you think about those folks, they can get started much more quickly now. They can borrow an expert from AI, and they could learn the things that they didn’t have access to before. That’ll be a great democratizer. I think it’ll lead some people to scale quicker, but also lead people to say, hey, I’m going to do something else with my time. And I think that that can work.

REID:
In terms of the lens of what you’re seeing—there’s a huge amount of potential in this, right? And, you know, I think every Silicon Valley, you know, startup entrepreneur, you know, is maxing AI—you know, token maxing and so forth. Are you seeing this more broadly across the country and across the world? Is there more than just the massive potential and the promise, but actually, in fact, growth in this kind of entrepreneurship?

RONNIE:
I do see that. One thing you’ll see a lot of is in other parts of the world where they might not have as much access to, let’s say, venture capital. Trying to get started off from zero to one is very difficult. You’re seeing people get started more quickly. We released a report six months back now showing that there are probably 4 million Americans who are using ChatGPT in some part of the entrepreneurial process. So it’s happening here as well. I think the key thing is brainstorming, ideation, and then going from step one to the next step after that. I wrote a paper a long time ago on the entrepreneurial process, and this paper documented that the most common failure point is just telling someone else about your idea and getting feedback. Sounds crazy, right?

Most of us, especially if you’re raised out here, you say, how could that be a hard point of friction? But it is for most people. And imagine now overcoming that friction—being able to ask AI, hey, what do you think about this idea? And getting constructive feedback. That’s where I think it can really matter. For small businesses already in operation, they can optimize different parts of their product portfolio or their HR strategy. A restaurant in Southeast Asia, for example, now can figure out, okay, what should be on the menu? What sells the best? Is this location the right one? These are all things that you couldn’t necessarily hire a consultant to do at the prices that they’re charging. But now you have AI helping and enabling it.

It’s this AI complementing entrepreneurship. I think that’s really, really important as we go forward. And it’ll be a big part of the economic gain as well.

ARIA:
So going back to those productivity numbers, you know, everyone talks about the advent of the internet, and, you know, in the ’90s, we didn’t see productivity for a long time after the IT spend. But Jason Furman, Obama’s former economist, he says that the productivity gains are already in the numbers. Anthropic has sort of shaded that direction. You said it’s going to be another 18 to 24 months. Why do you read the data differently than others, perhaps?

RONNIE:
I think that it is already showing up in GDP, so we should be honest about that. If you think about capex, anywhere from 40 basis points to 1% you’ve seen of GDP growth can be attributed to the infrastructure investments. That’s a big deal. And that’s something that’s been giving good uplift to the US economy. The next part is translating into productivity and what economists call TFP. I think when you look at that—I’ve seen some of the data Jason’s citing—it is consistent with it showing up in the productivity statistics. I don’t think we yet know for sure that it is AI. We’ll have to do more work to figure that out. And to your point, we won’t be able to do a lot of this work until it’s already in the past.

In 1995, there’s this famous Fed meeting where Greenspan and Yellen are talking about this, and Greenspan is convinced that he sees it. And Yellen—who’s an amazing economist and a Berkeley professor, so a lot of respect for her—she was saying, don’t yet see it in the data. Ends up they had to revise those numbers from that meeting way upwards. And so I do think we could be at the beginning of that moment. For me, though, what it means to show up and ascribe AI as the cause—that will take a little bit longer. But all of us—I mean, we can’t be fooled by our own eyes for that much longer, right? We’re seeing AI have a lot of beneficial impacts in businesses. Enterprise leaders are doubling down. They’re using AI across the organization.

I think—I believe in markets and incentives enough to know they must be seeing a payoff from that. And that’s going to translate into broader productivity. Not for all firms all at once, not for every part of the world. But I’d say the future is already here; it’s just unevenly distributed, to paraphrase William Gibson.

REID:
Well, actually, by the way, the funny thing is Gibson told a friend of mine he never said that quote, but he likes it.

ARIA:
He’s happy to have this.

RONNIE:
Maybe I should just say—

REID:
Just unevenly distributed. He’s like, it’s a great quote. Someone should say—

RONNIE:
That’s how—William Gibson. Thank you for that. That’s important.

REID:
It’s a really important quote.

RONNIE:
It’s a very important quote. The best thing he never said. That’s amazing. But that is where I feel AI is today. And if you look at where firms are adopting it, there’s a lot of uneven adoption, because some firms are maxing out power users on this, and other firms are—you know, we call these the median adopters—they’re far off that. And that spread is just growing and growing. It has big implications for how AI will translate into the economy.

ARIA:
Can I ask you—and I don’t know if you studied this outside the US—but you said it’s unevenly distributed. Are there pockets around the world where you see greater AI uptake than others?

RONNIE:
I think if you—I’ve gone to 12 countries since I joined OpenAI, so it’s been an amazing opportunity to go around the world and talk to enterprises, but also talk to just regular people, government officials, about how they’re seeing AI. Wide dispersion of views. I think in East Asia—in Japan and South Korea, and I spent a lot of time in those countries also before, in my academic and government career—a lot of optimism around AI and its ability to transform their economy. I think it’s driven by previous success with technology, demographic differences between those nations and the United States and Europe, and also this idea that many of the companies in those areas are already benefiting.

If you look at the chip supply chain—you look at South Korea’s economy with SK hynix and Samsung—you see those impacts. So I think a lot of optimism around AI, a lot of heavy enterprise use in those areas, and a lot of firms that are closer to the frontier because they’ve already adopted other technologies and now they’re using AI. I think in other countries, it’s more of a consumer story, where you’re seeing consumers harness the value of AI, the enterprise sector less well developed. But to me, that’s a matter of time and very consistent with a lot of diffusion patterns of technology in the past.

REID:
So actually, I’d be curious: what would be top three positive heavy users of AI in other countries, and top three negative surprises—like, you’d think they would be, but they aren’t?

RONNIE:
Oh, I think that’s interesting. I mean, on the positive surprises, I think countries like Indonesia and Brazil show—really, and we’re going to release some data to this effect, too—much higher usage than you might expect. And a lot of it probably is driven by the youth of their population as much as sort of things like education, income levels. You know, AI is really used heavily by people 18 to 25, and that’s a pattern that’s been seen all around the world. And if you look at countries with young populations—I had the chance to go to both Indonesia and Brazil, and I remember visiting universities there, and wow. And I know you guys—I spoke at a lot of these talks, but it was amazing to see the audience that turned out.

And what was so interesting, too—I remember this in Jakarta specifically. I’d been there as a kid, and now going back, the conversation in Jakarta was not that far off from the conversation in San Francisco, because these students were using AI the same way that their colleagues in San Francisco were, because it’s sort of a—you know, they can get these ideas and figure out these skills and do these things much more quickly. There was no lag. That was interesting to me. Those are probably some positive surprises. India is the other one, which, you know, many people wouldn’t call a surprise anymore. But when I went to the AI summit in India, it was just stunning to see how many companies are being built around AI.

Also, some of the bigger conglomerates there—their sort of AI strategies. One thing that’s important to remember is there’ll be a lot of startups, but incumbents have gotten the memo. They’ve read all the HBS case studies. They don’t want to be the dinosaur in those case studies anymore, right? They don’t want to be the one disrupted. They want to make sure they’re adopting AI. Negative surprises, I think—I’ll only say, I think in different parts of the world there are more barriers to AI adoption, right? And so I think in Europe right now, they’re struggling with, how do we use AI productively in the enterprise? How do we maintain sort of control over the key sort of parts of the stack that matter to us?

How do we do that in an economy that is, yes, globalized, but also where the geopolitics are shifting? And so I think in Europe, evidenced by the Draghi memo and other documents that have come out, there’s a lively debate about that. So I’d call it less of a negative, but more of, like, a real active debate over AI’s role in enterprise and society in Europe.

REID:
Yeah. So the IMF says that the capex boom is propping up US growth before any productivity shows up. And so two macro forces at once: a capex boom pushing up demand and inflation now, and productivity that should pull inflation down later. So over the next two years, how does this play out? Which one wins—if one wins, et cetera?

RONNIE:
Economists love a good horse race, right? So, I mean, okay, on inflation—multifaceted, right? I mean, if you see, like, the components driving inflation today, whether it’s geopolitics, the price of groceries, some of the things that have nothing to do with AI—the component that’s related to—I might be looking at the supply chain for hardware. I mean, you look at DRAM prices, something I paid a lot of attention to when I was in government, and you look at the ramp that they’ve had over the last year—it is enormous. I think the part of inflation that’s driven by these supply chain bottlenecks, which I think is a small part, ultimately—that part, I think, should recede over time as we build out more infrastructure and get more supply.

It does depend a little bit on the competitive strategies of the companies involved in parts of the value chain. I think AI over time has the opportunity to bring the price of a lot of goods and services down. That’s the force I expect to win out. Two years might be a little quicker than I might see that happen, but I think over time that effect’s going to be larger. In San Francisco, they call it the Jevons paradox. I just call it downward-sloping demand curves. At the end of the day, as you make the price of doing something cheaper, people might demand more of it, depending on if elasticity is a certain level. So I actually think we can see an expansion of a lot of markets with AI. And I’m seeing that in some of my own work already.

REID:
So far in our conversation, we’ve been reading indicators in terms of the whole picture. Now assume AI keeps delivering, and we jump to this far future of 2031.

RONNIE:
Oh my gosh, that is the far future.

REID:
Exactly. Describe that economy in detail: how fast it’s growing, who’s working where, and what a paycheck might look like.

RONNIE:
Sure. I think—look, if you look at AI adoption patterns, countries like the United States that are ahead on adoption will have reaped more benefits from AI if the scenario obtains. So 2031, you’d expect to see a repeat, maybe at a higher magnitude, of what you saw with the IT revolution. There’s this famous paper by Stanford economist Nick Bloom, John Van Reenen, and Raffaella Sadun looking at IT adoption by US and UK firms during that critical period we talked about. US firms adopted more; they had complementary management practices. That’s really important. And they got more productivity out of it. That productivity legacy really drove a lot of growth in the US. You’ll be seeing that in 2031. If AI pays off the way we think it will, firms in the US will have adopted AI.

They’ll be going global with it, and the US growth rates will be increased. Now, as we talked about earlier, I think if you get 1 more percentage point, 2 more percentage points—3 to 4% growth—that would be really significant. And in steady state, those are the kinds of things to look for. Sector-wise—I’ve written about this a little bit, something I feel strongly about—I think education and healthcare are two places that you’ll see a lot more employment. Why do I say this? As societies get richer, like in the scenario you articulate, we tend to spend more on education and healthcare—the things that really matter to us. Our priorities come out when we have this bounty from AI.

And I think taking care of people and educating them is a big part of it. When you add technology to those sectors, you end up actually usually adding a lot of people as well. That’s been the legacy of educational technology and a lot of healthcare technology. I also think, preference-wise, if you ask me whether I’d rather replace a kindergarten teacher with a robot or lower class size and have two teachers for half the size of students, I’d much rather hire more teachers. And I think most Americans would make that choice. I think they’d make that choice with doctors and nurses and other caregivers too. So I think you’ll see a lot more employment in those sectors. The big question about paychecks will be, do we make those jobs better?

Because right now, if you shift an economy from where we are now to more eds and meds, which are about one out of seven jobs in the US economy—if those jobs pay less than the jobs people came from, it won’t be as beneficial. But those jobs need to improve and need to be rewarded. There are lots of ways to think about doing that.

REID:
And I know it’s not as major for hiring and employment, but would entertainment also be one?

RONNIE:
I think so. You know how many people have a kid who has told them that she wants to be an influencer when she grows up? Right. I always joke, like, economists love to talk about all the new job titles that exist, and usually we’re talking about cybersecurity, but maybe it’s influencers nowadays. So yeah, I mean, I think entertainment is going to be amazing. Podcaster—we didn’t even have a word, you know, back in the day. So I do think that these new categories of being able to use AI and other tools to provide entertainment—it could be sort of disseminating this on platforms like YouTube, but it could be different mediums that we haven’t thought about, creating videos and art. There’ll be a lot of opportunity for this.

The question will be, how does the demand channel work? How much AI can people really consume? Will it be consolidated among a certain amount of players? These are all things to watch as you go forward. How democratized will the demand and the supply be? That’s, I think, the thing to watch, but I think entertainment could be a big, big one.

ARIA:
All right, you basically guessed my next question. I’m actually—I do want to talk about sort of hiring more teachers, because everyone wants small class sizes. But all the research I’ve seen shows that a better teacher is much better than smaller class sizes. And so sometimes, how do we use AI to get us better teachers as opposed to just more? And obviously the—the how much we’re paying them really matters there.

RONNIE:
Of course.

ARIA:
But you hinted at the care economy. We talk about that all the way—all the time here. We talk—we talked to Anne-Marie Slaughter about it, Spring big ventures, insurers. It’s something that feels distant from AI, but actually can have huge implications. And so you’ve argued that in the AGI economy, it’ll actually expand the care economy rather than shrink it. More people can be nurses, doctors, teachers, et cetera. But where does that money come from? Who is going to pay for all of those people who are entering the care economy? How do we make that happen?

RONNIE:
To me, if AI pays off in the way that Reid was asking me to consider in that scenario, we will have the funds to do that. Now, part of that might involve changing the way we tax labor and capital. Okay? But if we can basically get the economics and the incentives right and the government channel right, we’ll be able to channel the productivity bounty from AI into these sectors we care about. And if we don’t do that, then we need to set up better institutions to do so. But to me, this is what government’s primary role is, right? We raise more tax revenue through this sort of amazing AI dividend, and we’re going to channel that to the places that we as a society want to spend on.

And I would say, evidenced by previous history, Americans really want to spend on education and healthcare. There are differences, of course, on how much to spend, efficiency, but in general, we direct lots of resources to these two sectors. And I think we’d continue to do so, particularly if we had, like, a big sort of productivity boost. So that’s where I think it’s happening. It may require changes to the tax system and the way we distribute funds, but if AI doesn’t deliver that bounty, right, we won’t have that opportunity to do it. So I think in a world where AI is really successful and we adopt it and we see benefits, then we’ll have that money to invest in education, healthcare. If we don’t, the economy will change less than in the citizen.

ARIA:
I also think it’s so critical that AI not only produces those funds to help with healthcare and education—but anyone will tell you, we want to spend more there, but we have to have better services. We need better healthcare; we need better education. And so if we can use AI to also make those services better and more plentiful—I mean—

RONNIE:
I’m with you. I was thinking this, Aria. When I was filling out financial aid when I was a kid, we had the FAFSA form. I don’t know if they still—I remember it, right? My kids aren’t at that stage yet, but I imagine they might still have something like that. And I remember the whole talk when I was a kid was the internet was going to provide you information. So the internet would give you the information to fill out the FAFSA form, but you still have to look up the information and put it in the FAFSA. I think now we have the potential for navigation, which is what people actually need, right?

If you think about people navigating a financial aid form, if you think about a kid in a low-income setting trying to navigate some sort of school voucher program, or even someone trying to get onto the right sort of organ transplant list, these are all issues where we’re trying to allocate something that is scarce and people are trying to get access to it. But imagine if your agent was doing that for you. I mean, so we talk a lot about all these great agentic use cases in shopping, which is fantastic. But what if agents could help you navigate bureaucracy? And on the other side, the World Bank, the IMF—all this aid that we distribute through lots of organizations could help countries develop agentic platforms of their own.

Now you’re allocating slots with agents who are hopefully going to get to an efficient market outcome if they’re governed by a set of rules decided by the democratic populace. But at the end of the day, that’s a huge opportunity. So I think navigation rather than information will be a big step up that can hopefully help with some of these issues. On efficiency of services—

ARIA:
Can I ask one more question? I’d say it’s super—like, when I think about making it easier to fill out forms, bureaucracy, et cetera. But everyone’s talking about now how previous technology and AI has actually reduced some of the friction that we needed, whether it was friction to not gamble or friction to do these other things. And some of it actually is also friction to financial aid, because we actually don’t have enough aid to give to everyone. And right now the bureaucracy helps us mete that out. Are we going to run into a problem where, with the reduced friction and increased efficiency, we don’t have enough money actually to serve all those people?

RONNIE:
It’s a great question. This is like the restaurant where everyone—everyone’s going for reservations with their agent, but the restaurant just has a phone, right? It just keeps ringing. But the restaurant needs an agentic platform too. This is the answer. So you need platforms on both sides that can talk to each other. Because right now it’s the same problem I have at work. If people put a lot of new AI-generated content on my desk, I still have to decide what I want to pass through and prioritize. But if you have a world where people are applying for slots or trying to get access to a scarce resource, but it’s also being mediated by an agent platform with a set of rules—

And we’ve done this in a lot of areas—then you could see tremendous welfare gains. And there are lots of economists who are working on this. This, to me, is like one of the most exciting things about AI, and one that people often overlook. I think we think a lot about AI undermining expertise, which is really important. But we don’t think about where AI is the first expert, if that makes sense, for a lot of people. If you’re not within 100 miles of a behavioral health specialist, you don’t have access to that. And so AI providing you someone to talk to can be very useful as a first pass. It is true that there are lots of times where AI could have a negative impact on expertise as well.

We need to think about that in our markets and our government. But in a lot of places, people didn’t have access to begin with. And these are the things, I think, that we’re going to find. And it’ll be a big productivity channel as well.

REID:
One micro question on this that occurs is that there are obviously the cost curves—where we invest, how jobs are created. Medical and education, I think, are two great areas. It’s also that it becomes 24/7, and how does that time coefficient—like, on a services basis, it really matters. Like, for example, it’s 9 p.m., your kid has a medical condition, your grandparent has a medical condition. What do you do? Do you know—can you get to a doctor? Can you get to a—should you go to the hospital? Can you get there, et cetera? So how does this kind of 24/7 timing get reflected in the economic picture?

RONNIE:
Oh, the always-on agent is a big question in terms of how it’s gonna affect the economy. One way is, you’re right, for some of these tasks where consistent effort without sleeping is really important, you’re going to see tremendous benefits from unlocking a persistent agent. That’s something we’re studying right now. But in other areas, right, the agent might need to ask for clarification or permission, and it still might matter whether the rest of the world is up. This is another thing. The second thing is we have to prevent—going back to Aria’s point—overloading the system if it’s not ready. Persistent agents can be great, but what if it’s overwhelming the system that can’t take them? We’re going to have to have rules of the game here.

Even as I work with my team inside OpenAI, we’re often trying to set rules around, okay, how much AI-generated content are we going to send to this Slack channel? Who is going to be responsible for looking at this code or this narrative? And if you’re sending something to someone that’s incomprehensible, it’s almost a tax on them. And so setting those boundaries—humans setting those boundaries—is really, really important. I think we’ll probably have to have some of this mediating the agentic economy. There’ll probably be a bunch of stories first that kind of lead us to think about this more carefully. That’s where I think we’re going to need it. I think more broadly—like, I’m seeing this inside OpenAI—I mean, you know, agents are working, you know, three times as many hours as humans.

We released a report on research acceleration. It’s one of my most interesting reports right now, looking at how AI helps AI researchers, but also helps people across the company. And if you think about three times the human hours, we’re going to open up a lot of opportunities here. There has to be someone else on the—on the other side to catch the ball the pitchers are throwing, though.

REID:
Yeah, well, it’s not only that—you know, because I’m reminded of, like, the thinking of The Mythical Man-Month. But also it’s also the mythical AI hour, because not only is it, you know, obviously 24/7, but it’s also parallelizable.

RONNIE:
That’s right.

REID:
So you get this kind of mythical AI hour or something that is—

RONNIE:
It’s like a little bit—it kind of hurts your brain after something. And by the way, I think that parallelization of work is a key underpinning to the hardware piece you talked about, which is, are all those hardware investments going to pay off? I think if you’re thinking about one person, like Aria, Reid, Ronnie, using AI at one time, like a chatbot—if that’s your world, it’s very hard to imagine the amount of questions I could ask that could demand that kind of hardware investment. But imagine these parallel agents working persistently doing these tasks. Imagine the highest, best use cases in cybersecurity, in financial markets. Then you start to see a case for where we might need this much compute to deliver a lot of value across the economy.

That, to me, is the thing maybe people are sleeping on a little bit outside of the AI industry.

REID:
Let’s shift to one of your economic colleagues. Nobel laureate economist Daron Acemoglu has argued for years that the direction of technology is a choice. So usage data from, I think, both OpenAI and Anthropic say that most enterprise use is automation, and that’s what customers ask for. Can labs choose augmentation, or can they slant in that direction? Obviously there are some very positive social circumstances doing—

RONNIE:
Yeah, I think one lab can’t do it on their own. I think you also have to think about government policy in this area too. From a lab perspective, I think it’s interesting to me to contrast the CEOs who are talking about cost cutting only versus ones that are thinking about new value creation, because the promise of innovation—and this is true of the internet, it was true of SaaS software, it’s true of electricity, the steam engine, and semiconductors—is it wasn’t just about reducing costs, making things more efficient. It was unlocking new products that didn’t exist before and creating entirely new industries that we didn’t have before. That’s also going to be possible with AI.

I think there is an opportunity there that maybe fewer CEOs are talking about but is going to be as important as the efficiency gains. I think we’ll see that. The second thing, though, I think is that as an organization, it might be hard to slant in one direction, because there might be other providers that are providing AI for efficiency gains only or advertising that piece. I think it’s got to be something that is involving labs and government thinking about this in a way that is going to be sort of sustainable for, in this case, the United States’ long-term competitiveness. And not having sort of these productive jobs in the economy could be a big harm. So we have to think about this carefully. I just don’t think one organization can do it on their own.

So I think I agree with Acemoglu. This is a choice. It’s a choice we have to make collectively. And he’s written about this in great detail around institutions. But getting together institutions that get us all together—that is the hard part. And we have to have trust, and we have to be able to execute in those institutions. And that’s something that we’re going to have to work on in the next decade.

REID:
And that collectivity actually would be broader than an industry or even a society.

RONNIE:
Of course, and there are geopolitical considerations here. It is interesting to me to see a critical mass of people in government, in the private sector, in the labs, in philanthropy gathering around these topics, though. And I think that’s an interesting sign that they’re thinking about this. But to your point, in a competitive global economy, these institutions can be hard to sustain. So that’s what we have to work on.

ARIA:
So you sort of were enumerating various technologies throughout history, and all of these technologies sort of spurred us forward, but they also created winners and losers. And you were saying that, you know, if GDP growth ticks up by 2%, 3%—oh my, that’s enormous, that compounding over time—or if unemployment ticks up by 3, 4, or 5%, that’s also enormous. I mean, we were talking to some folks last week when we were talking about the China shock, saying that that was really just 3 to 5 million jobs. And, you know, in a country of 350 million, that seems small, but it had huge implications, because it was concentrated pain and, you know, diffuse benefit to the millions of consumers who got cheap goods.

So if you look out five years, like, who do you think are some of those losers in the AI age, and how do you think that affects politics going forward?

RONNIE:
Well, first, I think you’re right that the shock doesn’t need to be large from a quantitative perspective, as a percentage of the population, to have deep impact on the economic psyche and certainly politics. I mean, I live in North Carolina, which was sort of adversely hit by the China shock. If you think about the way it shaped our state’s economy and politics for generations afterwards, it’s significant. And so you might be able—and David Autor, of course, has done great work on this—you might be able to look at that set of congressional districts or commuting zones that he highlights in red as affected by the China shock and say, wow, they’re actually a pretty small percentage of the total amount of counties or commuting zones or congressional districts.

But in the end, they could be really influential, particularly in, you know, a two-party system with the Electoral College—50-50, right? In a 50-50 country. So I think it really matters. And so I think with AI, we could expect a similar thing, where the impacts might be felt in particular industries, particular regions. And it could still have a huge impact on our political economy. I think to the extent that it creates winners and losers, the winners will probably be more diffuse, to your point, exactly as they were during the era that we were talking about—the China shock, the march towards globalization. I’m already seeing this in the data. Like, one of the big things about AI is consumer surplus. Economists like to talk about it. It’s kind of incomprehensible to anyone else.

The best way I can explain it is, like, if I use AI to plan a vacation, I didn’t pay ChatGPT money for that. I’m on the free version. I plan a better vacation, my family’s happier, we have a great time, we have great pictures. That created benefit for me, but that’s not counted in GDP. But if you said, hey, Ronnie, after that vacation, I’m going to pay you basically to give up ChatGPT for your next vacation, I’d say you got to pay me a lot of money. We had a great time. There’s value there. That value is called consumer surplus. And it’s about $300 billion a year or something like that.

If you look at AI and what it’s doing to the economy, that piece is going to create all these diffuse winners all around the country, all around the world. I will say, in places with less resources, it’s even a higher benefit. Because, you know what? Unlike a lot of technologies, the quality that they’re getting from their answers to ChatGPT is pretty close to the ones I’m getting, by design. And because they have a lower level of services on average in some of these key areas, the welfare benefits, let’s say, in India or Brazil, are higher than what I’m going to get in Durham, North Carolina. That’s a big difference. We’ll see these benefits diffused around the world. In terms of the negative effects, they’ll be more concentrated.

And I think that’s going to create and will contribute to existing skepticism around AI, which you’re seeing now. And I think that’s something that the political system is dealing with now. If you look at sort of the, you know, town-by-town debates over data centers, if you look at how central AI is in the 2026 election—I imagine 2028—our system is dealing with this in the best way it can. I prefer this to a system where we weren’t dealing with it and we had some centralized authority that didn’t want to address it. So I feel like this is the right way to do it. It’s definitely messy, but this is the way that we work through it in a democracy. So for me, that’s kind of what I expect going forward.

And I think you will see those things. And I think that—I’m confident that our system can withstand it. But this is where what we talked about before—setting up the institutions in government and out to make sure that we’re targeting the health of people who need it—is really important. And with the China shock, we were too little and too slow.

ARIA:
Well, that’s what I was going to say, is that, like you said, we’re already seeing this sort of backlash. Are there specific things that you think the government has to do? You obviously think we should continue to invest in AI, and it can be a huge advantage for the US or for folks around the world who use it. And we’re sort of in danger of not capturing some of that innovation. Are there specific things the government should be doing so that our citizenry can be more excited, genuinely, about AI, because it’s actually helping them?

RONNIE:
Yeah. First I’ll say—and let me start from the personal side. I mean, we have to empathize. And I empathize with each person who feels this way. It’s one thing—I mean, you might believe in something, you might be using it every day. You might think, on net, it’s going to be good for the world. But people feel differently, and you have to sit with that. And I talk to students all the time who have conflicted feelings about AI, using it. They have conflicted feelings about how it might affect their job. And—

ARIA:
They might be right. And—

RONNIE:
They might be right. They might be right. But even if they’re wrong, it doesn’t mean that their feeling is incorrect. So I kind of feel like I can be the economist who comes in and says, well, it turns out the unemployment rate is actually only 4.1%. No—first, you empathize. It’s a tough time to be looking for a job. The economy—the economy is changing really rapidly. The stories in the Wall Street Journal might be anecdotes, but they’re still salient. I was reading this one yesterday about a chemical engineer who’s looking for a job that made a lot of news. And again, these are things that are very salient. So first is empathy. The next you do is provide evidence.

You try to tell people, okay, here’s where AI is creating jobs. Here’s where it’s having a negative effect. You’ve seen this work out of Stanford that’s really well done, that shows the softness in the early-career hiring. The authors are very upfront about what they can actually say about that, but I think that’s the closest we have to sort of a negative effect. But present that data, let people interpret it, and try to direct them, if you can, to places that are growing. Healthcare jobs are growing. They’re not necessarily the same jobs that people might want if they’re working in tech, but they are growing. And try to give people that honest assessment of it. So that’s step one. You got to build trust.

Because in the end, right now there’s a trust deficit when it comes to AI, and there’s a lack of trust in a lot of different areas. But AI, I think, is an exemplar. Two is, I think the government has to—and government policymakers in a bipartisan way have to—assure people that, like, what I call our automatic stabilizers actually work. There are a lot of discussions about inventing new stuff, which I totally understand, right? But I just don’t want us to forget the old stuff, right? When I was there during the financial crisis or the aftermath of COVID, we had automatic stabilizers. When the unemployment rate goes up, food stamps go up, we might pay out more, and other kinds of benefits. These things also happen at the state level.

There are a lot of boring, non-sexy things we can do to fix the state unemployment systems. We should do all those things. And I think if we tick down that list, we’re going to find that we’re better prepared for whatever comes. These are the kind of, like, no-regrets kind of decisions. If unemployment doesn’t increase because of AI, we shored up our safety net for whatever else might happen. And if it does, then we’ll make sure we’re in a good place. That will also build trust. And the last piece is—I do think this is rhetorical, but also probably instantiated through policy—people feel like humans are in charge.

When I look across folks who are cynical about AI—and they’re very honest with me about it—there are folks who feel that their expertise and their professions are being threatened. The mathematicians are going through this discussion right now. It’s a really poignant and important discussion, I think, for us to have. There are other young people who feel like it’ll be hard to gain expertise and they can’t get their foothold in the job market. That’s another group. And there are other people who are worried about sort of the externalities created by AI on the data center side.

With each of these folks, right, you’re going to have to basically explain, you know, what are we going to do about all these externalities, all these negative issues? And the government has to basically convince people that, look, we can contribute to the solution. We can help you through these transitions. That’s really hard to do. But if humans are in the loop and in the driver’s seat, I think it’s going to give a lot more confidence. Right now, people sort of feel like things are happening to them, right? And they’re not getting to have a choice in a decision. It’s really important to make sure people feel like that. These are not easy things. I don’t envy our leaders and our policymakers.

I do think there’s no more important problem to work on, though, than this.

REID:
So you’ve been in government.

RONNIE:
Yeah.

REID:
If you wanted to shape how AI reaches workers, small businesses—you know, helps them—which office would you want to sit in for the next three years?

RONNIE:
I really—I remember in 2020 when Gina Raimondo called me, and she was then the governor—governor of Rhode Island. And I didn’t really know why she was calling me. But then right when she called me, my wife had texted me and said, oh, she just—there’s a rumor she’s going to get tapped to be Secretary of Commerce. And I said, Commerce? I said, okay. And so I talked to Gina, and Gina was really excited about me coming to work for her in Commerce. And I wasn’t necessarily that sold. I didn’t know as much about the Commerce Department. I had come from working in the White House in the previous administration. I knew a lot of people who worked in Treasury. I hadn’t thought about—

But then Gina kind of rattled off, as she’s really good at, like, a bunch of things that Commerce could do. And not all those things were things they had done before, but they could do. She started off with NIST, the National Institute of Standards and Technology. I’d known a little bit about it, but honestly, I didn’t focus too much on it. And she said, you know, who’s going to set the standards of all these deep technologies going forward—in quantum, in AI, in energy? It’s going to be NIST, this office in the Commerce Department. You know, all the issues around climate and sustainability—well, NOAA, right, is in the Commerce Department. All these questions about how to develop regions and cities—the Economic Development Administration.

She rattled off all these different things. And this was even before Commerce had gotten involved in the CHIPS debate—the CHIPS Act, export controls, all these other things. She told me a bunch of different things. I said, oh my gosh, Commerce is going to be the engine of a new kind of economic policy that has really been durable and bipartisan, I should say. The way I came up, a lot of this stuff was done in Treasury and the White House. I think if I look at what’s happened in the last 10 years, Commerce has become a major player, maybe the major player, on economic policy. And if your question is about firms and adoption and small business, I’d say Commerce would be the place to go.

So Gina would love to hear me say that, but I also think it’s true in this case, and it’s something that she taught me and remains true today.

REID:
Gina is great.

ARIA:
She was on this podcast. She ran a masterclass episode.

RONNIE:
I was hoping you asked her who her favorite economist was. Otherwise, make sure you ask her that, to make sure we were even.

ARIA:
All right. So building on that, access to frontier models is becoming a trade question in both directions. You ran CHIPS, as you said. So set the security piece aside. In five years, do you think that exporting American models grows the pie for the US, or is there—are we leaking the advantage when we do that?

RONNIE:
I think that Econ 101 always taught us, let’s just build stuff where it’s cheapest and has the most scale. And that’s how we ended up with the sort of center of semiconductor manufacturing at TSMC in Taiwan. Econ 102—and this is the one that people didn’t always take, right—is that externalities get created when you concentrate production in one geographic area, especially if it’s in a part of the world that could be prone to conflict. Externalities get created relative to the environment and sustainability. Externalities get created relevant to national security. So with Econ 102 in mind, you have to make those decisions based on thinking about those externalities and the ability on national competitiveness and so forth.

For me, I think access to AI models and intelligence—I mean, that’s going to be important around the world. When it comes to frontier capabilities in areas like, for example, cyber, I think there’s going to be a strong national interest for countries to be much more careful about how that capability is distributed, because the balance of power and security hangs right on these kinds of things. So I think, for me, those two things actually can’t be separated. They have to be part of the economic policymaker’s toolkit. One of the things I learned in my second tour of duty in government—in the first tour, I kind of worked on the economy, and I wasn’t really interacting with national security people or defense people at all.

The second time around, almost all of our work was what you might call geoeconomics—economics at the intersection between national security and what might be called traditional economic policy. I will say that is a sea change, one that really none of our generation was trained for, the way we learned economics. I mean, I remember asking about industrial policy, like, my first day of graduate school, and someone said, that’s what they do in Europe. Don’t focus on that here. But by the time I got back to the administration, we were thinking about, how do you create a vibrant semiconductor manufacturing industry in the United States of America? And those questions, I believe, are going to last beyond any particular administration. We’ll be asking them with AI, with quantum, with energy and synthetic biology, and a whole host of other areas. I think it’s a sea change in terms of the way the economy was studied and policymaking was done before that. So that’s where I think it’s happening.

ARIA:
And so when you think about trade policy, what are the trade-offs that you think about between protecting our own people and open markets? I mean, this is the debate for all time. I mean, how do you think about that?

RONNIE:
What’s going to happen in many deep tech markets is unlike the internet era, and I think people haven’t clocked this enough. You just calculated TAM by the population of the world or something. Right now, especially if we have different stacks in parts of the world, your TAM is going to look different. And there are some technologies—I mean, you look at the difference between quantum computing and the internet. We were having a big push to try to get more people on the internet all around the world. With quantum, it really matters who’s first and who’s best. I think that’s going to change the total addressable market. It’s also going to change the strategy of these companies. And you see this in quantum; you see this in AI.

A lot of people are looking at the interactions between governments and companies and saying, why does that make sense? That’s different than what we had before. The technologies are different. Deep tech naturally is going to intersect much more with national security than the previous consumer tech that we had before. That means a different role for government. And it means, I think, that you’re going to have to make that trade-off sometimes between protecting national security and a bigger market. Different policymakers, different companies will push in different directions. But I think—I wouldn’t be surprised to see that be a consensus across multiple administrations going forward.

REID:
So let’s go inside OpenAI through your team, rather than just kind of the overall organization and all of the different assets and whatnot. So tech companies have had chief economists for a while now—you know, Amazon, Uber, Microsoft, Google. Why does a frontier lab need you—need them? And what changes inside—what changes inside the company?

RONNIE:
I hope Sam is listening, because I got to justify—I think, number one, I mean, the chief economist movement, I think, probably started in the early 2000s. I remember Hal Varian. I was a—I was a student at Berkeley when he was a professor, and when he went to join Google—I mean, we were all—we were using his textbook back then. And we were all sort of like, wow, this is really a really big move. It was big at its time. I think people—it was unusual for someone like that to go take a role like this. And we all realized that it was the beginning of a new era. And you saw that with other economists at the places you mentioned, with the frontier labs. I think there are two things.

One is, there’s so much data that needs to be interpreted and understood to give society the tools they need to make decisions. When I think about the most important thing I do all the time, it is basically the idea of putting the data and insights out there so people at the Department of Commerce, in the sort of Bank of England in a different country, can understand what’s actually happening. That is my number one job. And I would say, with the exception of maybe a couple other organizations in the world—and on the consumer side, I think it is us, to be honest—we have the best data on this. So it’s a tremendous responsibility to be able to procure that data and put it out there without an angle, without a slant. That’s really important.

I think you need economists in-house who are going to do that. We have an amazing data science team. Data scientists are often thinking about the product decisions that are going to inform enterprise decision-makers. Very important. But to have an economist thinking about the big picture outside the lab is really key. That’s one. Two is, I mean, we’re creating economic actors inside the agent economy. When you think about what my MBA students and another generation are going to be doing, they’re going to be managing more agents than they are people. How are those agents going to behave? What kinds of economic organization will they support? What incentives will govern their activities? These are all things that economists actually have a lot of experience answering.

So I feel like the combination of the data piece and applying economic reasoning and theory to this agentic economy are two reasons why, you know, frontier labs need a chief economist.

ARIA:
Well, also, you’re unique. You publish with outside academics and go through peer review, which is unusual for a company. So what do you think that you can see from your vantage point inside OpenAI versus a perch at a university? But then also, what do you think it costs in credibility when people might say, well, you’re at OpenAI, why should I even trust what you’re putting out there?

RONNIE:
It’s a good question. And Aria, I thought about this a lot in 2024, when I decided to take the job. I knew that my life would never exactly be the same, because working in government, I had had that hat. This is the first time I had gone full-time since graduate school to a private sector organization. And I really realized immediately, hey, I gotta work with outside academics. One is, there are gonna be a lot of people who see a lot of things that I don’t see. So if I can help them, under the right sort of governance and security regime, get a window into what we’re doing, they can teach me things with their insights. And I’m not gonna have all the expertise on my team.

I’m not gonna have a theorist or a macroeconomist or all the things that I might want. So working with someone else is really important. Two is, let’s publish. Economists talk to each other through working papers and publications. It’s old-fashioned, but if you want to influence the way economists think about AI, you got to write these things. And it takes forever. I mean, I am under review in a bunch of different journals, and it takes forever. And sometimes I wish I could plead with the anonymous reviewers, please, I don’t have to do this. But it’s really important. And we go through those aches and pains because once it gets published—these papers—we’re going to build the body of knowledge to influence how economists think. That’s how I learned as well.

So we have a paper right now—it’s a working paper. It has, like, 850 citations. That’s amazing. In sort of a short amount of time, less than a year, it means 850 people saw good reason to build on that work. If I can put that out and get it published, hopefully more people will cite it. That’s just the working paper. That’s why you write; that’s why you put this work out. The credibility thing is always challenging. It was a hard, hard line to follow. Here are the two things I do. One is, in addition to working with external collaborators, I try to give people as much data as I can for them to work with, to say, look, if you want to run the regressions or do the analysis, you can do it.

And that way it creates transparency. Two is, the journals themselves are peer-reviewed. So hopefully if I get the papers in journals, people will say, well, at least someone outside of OpenAI thought that that was good. And the third thing is, you got to seed other people’s research and not do it all yourself. So we have a program called Exchange. We funded 14 different teams so far. They’re doing research with our data. They can be independent of OpenAI. Those are the ways to do it.

ARIA:
Awesome.

REID:
So, as I understand it, your research concentrates—OpenAI concentrates—on how people and firms are using AI right now. Why did you choose that as a focus? And then what do you give up in order to keep that focus?

RONNIE:
Excellent. I mean, it’s interesting. When you start a team, you have to decide what your priorities are. And that was really hard. I didn’t want to give up anything. You’re like a kid in the candy store. But actually, this is where a lot of my colleagues and friends at OpenAI who had a lot of experience working in tech organizations—they said, Ronnie, you’ve got to choose. You got to prioritize. And I didn’t have all the knowledge I needed when I started, but I did listen to smart people. And so I said, look, I’m going to choose a few things I’m going to do, and when it comes to placing bets, I’m going to place bets on the here and now and the data rather than, let’s say, making predictions in the long, long term.

Why did I do that? One is, I felt like I had worked in government, I’d been in the seats, and I thought if the objective here is to give policymakers and enterprise leaders, like, actual data they can use, it’s better to tell them about the now—which they don’t have data on, by the way—rather than making predictions, predictions that you might have to update every six months and say, actually, let me push that back another year. Oh, actually—I just thought it’d be better to do it that way. There are trade-offs when you focus on the here and now. The results you come up with are sort of like, oh, I already knew that. And you say, well, you think you know that, but you didn’t already know that.

So I do think an empirical foundation, for me, was what I feel the most confident about. I feel like this is a very hard world to make predictions about. And I think while some of the predictions can move the Overton window and help people think about things that they haven’t thought about before, it can also create a lot of noise that makes it harder for enterprise leaders and company and government leaders to make decisions. And if I look at the most useful stuff we’ve given to policymakers in terms of what they’ve responded to, it’s often like, oh, you’ve told us how it’s affecting the job market now. That’s really important, because my job is now, right? That’s why I made that decision, and it does come with a trade-off.

ARIA:
So when you talk about credibility, I feel like there’s no—well, maybe not no bigger, but one big piece of credibility right now is, are you blue or are you red? And so OpenAI brought in Jason Furman and Michael Strain, economists sort of from both sides of the aisle, into the organization. I would imagine, to get that credibility in a bipartisan way—like, why did you bring them in, and what do you gain from these sort of external credibility points?

RONNIE:
I think you had to bring in people with a wide variety of perspectives, right? And in economics—I mean, the truth is that, you know, people who are associated with political parties, they probably agree on the economics 85% of the time anyway. But in other areas, it could be more diverse. So I think it’s important that people see themselves in OpenAI. They see that their perspective is being represented. Particularly because—you think about, there are a billion people a day who are going to talk to AI for advice about finance, about politics, about relationships. And if you’re building that sort of machine intelligence that’s going to provide that information, people have a vested interest in making sure that it’s unbiased and that it’s based on the best evidence. And people have different ideas about that.

So by bringing people in to consult or be advisors, you can build that trust with communities. I think it does go beyond that. You have to go out and meet folks and talk to them about this, right? And show them the data to say, here’s what a prompt is, here’s what AI produces. I think when you do this around the world, as I mentioned, you can build trust globally. We’ve also done it in the United States. I’ve been to so many states talking about AI. I’ve done AI workshops and boot camps. I used to be the person teaching, but I’m the person learning as it advances. That’s another piece. Showing up is a huge part of this.

I think we found this in all our professions, and in my previous life, when I was in public service, showing up at the place, I think, is a great way to defuse a lot of mistrust and tension. You know, if you spend your life on Twitter arguing with people, you forget that when you show up, people are like, wow, it’s a lot harder to be rude to that guy to his face. And I think I’ve learned that, like, going around. So that’s how we do it. And I think those two are an example of that.

REID:
So who’s the economist—economists—you’d most like to see argue with your data, and about what?

RONNIE:
I really—I think Daron Acemoglu has been articulating a lot of sort of counterpoints on AI that are really useful. Some people would say, oh, is that threatening? I say, no, no, that’s how you get better. And he’s written a lot of influential books, and I’ve learned a lot from his work. So for him, I’d love to share that data with him. And we’ve had conversations, and I’d love to see if he draws the same conclusions we do from it. And I think if I do this right, he would. He might differ on the interpretations or the long-term projections, but he’s the one I think would be great to interact with the data. And he’s done this kind of work before.

So that’s the person, sort of, like, currently that I’d probably want to interact with our data.

ARIA:
Awesome. And so part of your job is also measuring what AI can do for real work, not just how it scores on exams. And so GDPval is OpenAI’s test of that. And it’s built from tasks written and judged by experienced practitioners in 44 occupations. But how do you turn, like, an entire profession into a test? And what does that test then teach you about its claims on the economy?

RONNIE:
Yeah, I mean, let me start at the beginning, if I have a second, which is a lot of the—I think the interest in AI and jobs started from this idea of looking at tasks. And it was a great place to start, because if you ask me, hey, Ronnie, tell me what you do at work, and I can write it down—or better yet, ask the Department of Labor what I do at work, and they’ll write it down. That’s called O\*NET, right? So O\*NET is basically a list of tasks. And then you ask, okay, can AI do this thing? And machine learning people love benchmarks, right? So you’re able then to test the AI against these tasks and say, does it do it better than a human? Is it equal?

What do you do with ties, all that kind of stuff? And that’s essentially where the GDPval idea comes from. Originally, though, I think a lot of people looked at, let’s say, 10 tasks in Ronnie Chatterji’s job, and they said, if AI can do five of them better than Ronnie, he’s got, like, a 50% chance of being unemployed. That’s not how it works at all. Like, I just can’t be more emphatic that that is not how it works. I still remember my dad got a computer in his office in 1984. He ran big regressions and stuff. He’d go into, like, a big mainframe computer room with a punch card. He’d tell me all these stories. He got the computer—every single one of his tasks, as an economics professor—I’m a second-generation one—basically could be done by the computer. But he wasn’t out of work in 1984. He used the computer to run bigger regressions, more models, more work, write more papers. And so the idea that exposure to AI equals bad is actually not the case at all. In fact, sometimes exposure can be good if you can leverage AI for what you’re doing. The other thing people forget is tasks are, for a lot of us, endogenous to the job. It’s not like someone just says, Ronnie, you’re only going to do these 10 things. Ronnie, you have a high-level objective that’s a little bit fuzzy.

Here are the 10 things that most people do, and you figure out the things you want to do, and you realize, gosh, I can reach that objective in other ways. I’ll add tasks 11 and 12 and also drop one and two. That’s how a lot of us, particularly in white-collar work, work. So I think while it was really interesting to look at the task-based view and kind of machine learning benchmarks as a result of that, it led to a misleading assumption about how many jobs would be destroyed by AI, and it completely missed the idea that you could increase demand for those things. I would say that start kind of got us to a point where we’re making predictions that haven’t come true, right?

And done a disservice a lot of times to people who are trying to figure out what’s going on. However, when you look at GDPval, it’s really interesting to see how well it can build slides, it can build spreadsheets, it can do lots of these things in the profession. So I’d say for a financial analyst, you can now see how GDPval performs on different aspects of their data job. We try to write down, based on O\*NET and other classifications and deep research with those folks, what they do. But surely the job of a business analyst at McKinsey has changed a lot in the last three years. That’s the moving target that keeps us from thinking that, oh, we can do everything in someone’s job. Their jobs change. The Frontier Evals team that built GDPval—they deserve a ton of credit, though, because they’ve created—they’ve created kind of a benchmark way for people to keep track of AI’s economic impact. It’s enormously invaluable. We use it all the time. I think jobs also change, and that’s what we’re finding at the same time. And they might change faster even than AI.

ARIA:
Yeah.

REID:
One of the things—because I’m also familiar with O\*NET from LinkedIn.

RONNIE:
Oh, of course, yes.

REID:
Have you guys run any analysis about, like, how accurate O\*NET is, or updated in an AI age? Because it’s a little bit of a—old is not quite the word I would use, but it’s a bit dated.

RONNIE:
So, Reid and Aria, you have a general audience for your podcast. Unfortunately, this is going to take us down a rabbit hole. You’ll only have five listeners left by the end of it, because—but I will be short. Economists in general really want to improve O\*NET. Okay? I mean, it’s the best we got. It probably wasn’t fit for the purposes it’s being used for. It is alarmingly comprehensive, and so we’re able to use it for a lot of things that it probably wasn’t meant for. I think a fantastic project—and our team has been talking about this a lot; we submitted a comment around this to the Department of Labor—is we should build a new measurement infrastructure around work and AI, for sure.

This is the kind of wonky stuff that you can do in a job like mine, where, you know, people will listen. You have the capability and the credibility to do it. We would love to see a new investment in data infrastructure, because O\*NET is not fit for purpose for what we want to do. So more to say on that, but that’s, like, for a very nerdy podcast.

REID:
Yeah, well, but—and also super important, which is the reason I asked it, because I was like, I knew that from a pre-AI era with LinkedIn, and it was already creaky.

RONNIE:
Right. And that was applied to LinkedIn, right? And then we have so much better data we could build, I actually think. And that’s the kind of thing we should build measurement around at the US economy level and around the world, and imagine an international effort. Because one thing we haven’t talked about is the canaries in the coal mine might not be in the US job market. They might be in the Philippines call center market; they might be in the Indian BPO market. And I think looking for those can be just as important. And if we have the right kind of labor taxonomy, it’ll be more helpful.

REID:
So your team sits on more data about how people work with AI than maybe any organization in the world—certainly any statistical agency.

RONNIE:
Yes.

REID:
So how do you see—you already had some of the focus on the now and then truthful reporting of it, which you’ve covered before. But what do you see around your responsibilities, and how do you run the timing and the sequencing and the packaging of sharing that with the world, and what they get to see?

RONNIE:
Yeah, I mean, first of all, from the content marketing side, it’s really important to learn from the best. We actually have someone on our team who worked at a company called LinkedIn, who was doing this in terms of data insights and making sure—because LinkedIn had a lot of this early data and a lot of relevant stuff about the job market—and creating that in a way that people actually want to engage with it is really important. And it’s a crowded space, as you well know. A lot of people are writing about the future of work. So one is thinking about, okay, who is our audience, and what kinds of data and insights do they resonate with? And we’ve done, like, audience segmentation and think about who responds to our work.

That’s what platforms we want to use. I mean, going back, LinkedIn and X are different platforms. You know, Substack is a different platform. How do you think—think about what you’re putting in those different places? In terms of the substance, we focused a lot on thinking not just about jobs, but actual work. We’re thinking about, okay, we can count the job numbers the same way that they can at the Bureau of Labor Statistics, and they’re probably better positioned to do that. What we see that no one else sees is how people are using AI at work. So why don’t we create indicators and measures—and we do this in our Signals data and our enterprise Signals data—that basically show things that only we can show? I think that’s another key.

You got to have the channels, but you also have to have things that only you can produce. Otherwise, why are people going to come to you for it? The last piece is the hardest, which is a lot of this is wonky. Statistical economists love to talk to one another. How do you make it relevant? Someone at the Wall Street Journal or a podcast or a Substack will pick it up. And what is the line between making it simple, where it can be communicated factually, versus sometimes things get butchered? To maintain that credibility that Aria was asking about, you need to make sure that people get the message right. I spend a lot of time worrying about both.

Like, are we being too high-level and wonky when we communicate? Are we actually getting something, though, that is right? Because that is, like, the most important obligation. I think we talked about this in the green room too. It’s like, you know, I give a lot of talks where both sides are disappointed. You know, someone will come up to me at the end and be like, you didn’t go into reinforcement learning enough? And I’ll be like, ah, okay. Another group will be like, I just opened up ChatGPT for the first time. It’s like, how can you do that, right? Same thing with economic statistics. So we’re trying our best, and, you know, hopefully—hopefully it’s hitting the mark. So your listeners will let me know if it is.

REID:
Well, sure, it’s fascinating. So before we move to Rapid Fire, are there any questions that you had hoped we might ask you that we haven’t? Anything that’s kind of top of mind for you?

RONNIE:
I think the only thing—maybe what has surprised me most about joining OpenAI. And I would just answer by saying it’s nothing to do with the economics; it is really the capability. We talk a lot—and in my job, I talk a lot—about what people have got right and wrong about the economic predictions. And that’s a world I know well. I didn’t have much prior on the capabilities side of things. I heard a lot of stuff. I’m naturally skeptical because I’m an economist, but the increase in capabilities is stunning. And those who understood scaling laws and understood the direction of travel and the rate of travel—I mean, they were right. So I’ve learned from that.

I think you got to be honest about the things that you need to update on yourself and things that you maybe were right on. I think on the job market stuff and the growth stuff, I had it right. On capabilities, way off. And that’s interesting. And I think if you look at when I joined OpenAI, AI couldn’t do my daughter’s math homework. And now we’re solving problems that haven’t been solved by humans in 100 years. That’s in two years. You have to sit with that and be in awe. And it has so many consequences, but it’s also just completely mind-boggling to me. So that is what surprised me most about joining OpenAI—capability growth.

ARIA:
I think what’s also so interesting about that is—I think I just feel like there’s no other technology that changes so rapidly and that truly affects public opinion, because I see in my everyday life a lot of people I know still think ChatGPT and AI in general is bad at math. Oh, why would you talk to it about math? It’s terrible at it, remember? And so because that’s changing so rapidly, if they used ChatGPT 18 months ago, even, they have an outdated view. And they’re like, of course AI is not going to be good for me, for my job, for my life. And so we have to reintroduce the technology to people every month.

RONNIE:
I mean, we’ve been through multiple revolutions in this. When people tried it, maybe in 2024, it was in the chatbot era. It was an ask-me-answer machine. We then had reasoning models. I think people sleep on it now, but, like, the o1 series was pretty important in terms of thinking and also inference and optimization there. And now we’re into the agentic world. So people who tried it for math and said, oh, it doesn’t have an answer—they also are imagining, I think, asking, like, ChatGPT, can you solve this thing? And I think what they’re maybe not clocking is, you know, 10,000 agents working together collaboratively. This is like a research organization, not a chatbot.

And these are all things that I think we’re living in this world, but for a lot of folks, it’s moving so fast. And what we were calling AI in 2024 is certainly not worth describing in 2026.

ARIA:
All right, well, I’m excited—oh, you have something else. Go.

REID:
Well, should I ask you when you will have AI agents advancing the field of economics?

RONNIE:
Oh, you should. AI agents are already advancing economics. I think it’s in ways that are probably more pedestrian than people imagine if they’ve seen the math stuff. A lot of our work is—I don’t know if you should call it downstream, upstream; economists will be offended either way—but I think it’s downstream of a lot of the most theoretical math, in the sense it’s a little bit of art and science and social science. The art is, how do you identify the research question? There are some outstanding research questions that are sort of like the Millennium problems. But, you know, mostly I think a lot of economics is posing the right questions to ask. And so I think there’s that piece that is the domain of humans so far, in a really good way.

There’s also an elegance in how you answer the questions. We aren’t always exactly proving them in the same way. There is a theoretical branch, obviously, of economics where that’s more sort of the thing. But in a lot of empirical economics, it’s, like, framing the question the right way, placing the literature, and then using the data. You have to answer it in a way that’s very convincing. And I think AI is going to help with the ideation around what that idea could be, citing other literature that you might have missed, particularly in journals you might not read, which can be really important.

Connecting that research and conclusions to other people’s work that you might have missed, and probably making sure that you haven’t made basic coding errors, matching data incorrectly, and that the specifications you’ve run are actually robust—all the issues you need to worry about in econometrics—those are things I think are really important. One single prompt, write a paper—people on X disagree on that. I’m still writing my papers.

REID:
Yes. Yeah. So now we will turn to Rapid Fire.

ARIA:
Very special edition of Rapid Fire.

RONNIE:
Is there some music that goes along with it, or some sort of drum roll?

ARIA:
Ready? All right. What is the one number that you check every week that no one outside your building looks at?

RONNIE:
I’m super interested in revenue per employee at some companies I track. You know, when the paper on the job market came out of Stanford’s Digital Economy Lab, it was called Canaries in the Coal Mine. I thought it was a really great title, and I thought, what I really want are canaries in the coal mine for productivity. I want to know that—the Yellen and Greenspan debate, the debate you cited with Jason Furman. I want to know what to be looking at. And Jason and others and Erik Brynjolfsson have identified some of the numbers. I’m looking at revenue per employee, and I’m looking at a set of firms that I think are particularly AI-exposed, from things I know. And I want to see how that’s changing over time.

And that’s not always reported, let’s say, every month. But I am tracking that, and perhaps others are too. But that’s something I really care about.

REID:
Any surprises or interesting notes on that number?

RONNIE:
I think that the industries that you—you expect it might increase in, just to keep it close to my chest, they aren’t increasing as fast. And there are some industries that are sort of sleepers that are interesting. Yeah, that’s a—

REID:
Drum roll for the future.

RONNIE:
Exactly, exactly.

REID:
So the consensus among economists about AI that you think is wrong.

RONNIE:
Okay. I will say the path to AGI runs through the organization. I think economists and machine learning people usually miss this. Why? I think a lot of the machine learning folks are thinking about AI’s raw capabilities, and they’re thinking, well, if it can do all these—these things, like, it’s going to do them. They’re overlooking, I think, the role of organizations in mediating the adoption of new technologies. Economists have a branch of our field called organizational economics. But typically, I think, organizations are overlooked as a unit of analysis. I think they’re humans’ greatest invention, to be honest. And I also think that—I mean, it’s our ability to collectively pool our resources and talents and get things done.

And I think that we’re overlooking in a big way—economists and others—the role of organizations in making AI what it is. I think AGI—for it to, for us to get to this point, it’s got to work through the enterprise. That’s what I believe.

ARIA:
It’s the end of 2030. What do you think the unemployment numbers are in the US? Best guess?

RONNIE:
Yeah, I’m hoping that it will be 5% or lower. You know, I think that if AI pays off in the way we’ve been thinking about it, it’s not necessarily going to create a lot of disruption. I think if AI exceeds our wildest imaginations in terms of its adoption and somehow affects the job market in a very negative way, we’ll have to do the things we need to do to make sure we soften those transitions and get people into productive work. There are lots of policy levers along the way. I think between now and 2030, though, either we’re going to be able to deal with those issues or we’re going to find that AI’s impact on jobs is less than we might imagine. Yeah.

REID:
So the share of American workers using AI daily in 2030.

RONNIE:
Yeah, I think if it’s about 40% now—I think, according to a survey by David Deming and co-authors from YC at work—I think you’ll be up to 80% by that point. There might be some—I think this is interesting. Like, we’re using AI, but we don’t actually think of ourselves as using AI. So the number could be even higher, because I think we’ll be using AI embedded in a lot of other platforms.

ARIA:
The economist, living or dead, that you would hire tomorrow.

RONNIE:
I’m glad you gave the option there. I was having a conversation the other day which recalls us, and someone was just talking about the contrast between the way we learn economics versus a generation ago. And I have a comparable, because my dad did his PhD in economics in 1963, and one thing he told me was that he read a book for his class about John Kenneth Galbraith, or by John Kenneth Galbraith. And Galbraith isn’t someone that most current econ students would read, because it’s quite narrative. It involves economic history. The discussion I was having at this dinner was actually that, like, that would be very useful to a lot of economists today. Many of us don’t know as much about the history of our field as we should. I put myself in that category.

And so I think you need someone like Galbraith, who was the ambassador to India, as well as, like, a big public intellectual, to weave together the impact of technological change on economic development, on geopolitics, and domestic politics. And he was someone who just understood that at the highest level. So that’s what I’d want.

REID:
Awesome. So the one thing the economy can do in 2036 that it can’t do today because of AI that you would be willing to bet on.

RONNIE:
Oh, I think we’ll have agentic organizations, for sure. I mean, they will have the—

REID:
Entire organization.

RONNIE:
Yes, yes. I think that you look at what we can already use today with agents and how capable they are, then you look at what they can do in concert, working together. Really, the cost of compute is the main thing to keep an eye on to make these happen. Now, the questions on legal liability and who the ultimate residual claimant will be on the company are important topics, but I think functionally you’re able to have agentic organizations. They might be at smaller scales than the grandest ambitions, but I think that’s a safe estimate.

REID:
Great.

ARIA:
Ronnie, thank you so much.

RONNIE:
Thanks for having me.

REID:
Thank you for joining us.

RONNIE:
It’s great. Thank you.

ARIA:
Love it.

If you enjoyed this conversation, follow the show wherever you get your podcasts and consider leaving us a review. It really does help audiences find us.

REID:
And if you know someone who would benefit from this episode, send it to them. Thanks again for being with us on another episode of Possible.

Possible is produced by Palette Media. It’s hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Shaun Young. Possible is produced by Thanasi Dilos, Katie Sanders, Spencer Strasmore, Yimu Xiu, Aman Suri, Danny Garrison, Trent Barboza, and Tafadzwa Nemarundwe.

ARIA:
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil, and Ben Relles.

And a big thanks to Zach Herlick, Amanda Anderson, Josh King, Olivia Oran, Allie Mack, and the team at Midpen Media Center.