Scott Wallsten: On the morning we recorded this episode, Jason Furman typed one sentence into a system he built. Three hours later, he had a 90-page economics paper. I asked him what he did in the meantime.
Jason Furman: So I taught a class and had lunch.
Scott Wallsten: Hi, and welcome back to Two Think Minimum, the podcast of the Technology Policy Institute. Today is Wednesday, November 18th, 2026. I’m Scott Wallsten, president of TPI.
Even a year ago, writing an empirical economics paper took a trained economist, probably at least one research assistant, and several months. This summer, my guest built a system that does it in an afternoon. You type a topic, sometimes just one word. The system picks a question, finds the data, runs the analysis, writes the paper, and sends it through its own referees. He’s posted five of those papers so far, so anyone can judge them.
That raises questions for doing research, for training people to do research, and for policymakers. Congress and government agencies rely on studies. If every side in a debate can produce competent-looking studies overnight, how does anyone decide which ones to believe?
So my guest is Jason Furman. Jason is a professor of the practice of economic policy at Harvard, where he co-teaches the introductory economics course, and a senior fellow at the Peterson Institute for International Economics. He chaired the Council of Economic Advisers under President Obama. He also runs Econvitals, a site that tracks the economy and is written largely by AI. Jason, welcome.
Jason Furman: Good to be with you.
Scott Wallsten: Let’s start with that paper. Your system is called Paperwright, and this morning you asked it what the UK’s minimum wage increase did to jobs, incomes, and poverty. What happened between your one sentence and the finished paper?
Jason Furman: What it did is, first of all, I should say I first set the system up at the beginning of August, and it would take about six to nine hours, and it’s gotten faster and it’s gotten better.
So what does it do? It plans what the paper is. It does something that more economists should do. It pre-registers its research design. So it commits in advance: this is the data I’m going to look at, this is the outcome I am going to look at.
Scott Wallsten: I’ve noticed that, too. It’s a really big improvement over the way studies tend to be done otherwise.
Jason Furman: Oh yeah, it does. I mean, because you could imagine AI does the forking paths thing, where there’s a million choices you can make, and if you make all a million in a certain way, you can get to the answer that you want, or you can get to a really cool, exciting answer. Maybe that’s what you want, it’s just a cool, exciting answer. I would find these AI papers have often come out a little bit disappointing. And I think they come out a little bit disappointing because they’re actually honest in a way that researchers aren’t, because they pre-register the research design, they go out and do it, they find it’s not statistically significant, and they don’t manipulate every which way.
So now, to be clear, I think one could build a system that p-hacks, but for my sins, I’ve built a system that refuses to p-hack. But I’m not putting my name on any of these papers, so I don’t need to worry one way or another.
But anyway, it then dispatches different agents. Some of them read the literature. And the important part there is, to the extent possible, they download the papers and read them, rather than just guessing what they say. Another one pulls economic data, and it’s already pre-connected to about 125 different data sources, but often it finds new ones and figures out how to connect. Then there’s ones that draft the papers. There’s a very extensive fact-checking process. There’s several referees. There’s revisions in response to the referees.
And then the thing I like is at the end of it, you don’t just get a paper. You also get a memo that explains a little bit what it did, what some of the limitations were, and then it often has a few ways in which, if I wanted to be the co-author with Claude, I could make the paper much better. And almost always, the co-authoring with Claude, the thing it wants me to do is be an intern and go fetch some data that it wasn’t able to get because bots were blocked or whatever it was. But it’s basically like, yeah, if you go download this data, I’ll make the paper better, and I’ll offer you a credit.
Scott Wallsten: But there are opportunities for you to be involved in it along the way, too. I mean, in the case of Paperwright, the whole point is to not be. But according to its own memo, it asked a bunch of questions, but you didn’t answer them, and so it answered them itself. And again, it’s not a criticism, because that’s exactly what you’ve set it up to do. But what were those kinds of questions? What sorts of decisions did it make?
Jason Furman: I mean, it asked questions like, do you want to focus, for the minimum wage, on different ages? Or do you want to identify by different regions? Or do you want to identify by the rollout timing? Do you want to use this data set or that data set to do it, et cetera? And for almost all of those, I left it to its judgment.
I’ve also built into the system that there are two rounds of questions: ones that it comes up with after two or three minutes, and then, an hour or two later, as it discovers stuff, as it does even more work, it can ask another round of questions. But again, as you noted, those are both optional.
And there’s one paper I wrote that was just with a prompt, Triffin, T-R-I-F-F-I-N. And that was, I think, about a 70-page paper. I don’t remember the exact length. And it has a fully worked-out model that I think is actually a coherent and correct one. And a little bit like the minimum wage, it has some disappointing empirics that don’t quite live up to some of the hype that it itself did in its own abstract. And so I guess in that sense, it’s repeating at least part of what humans do.
Scott Wallsten: Have you thought about submitting anything to a conference or a journal?
Jason Furman: No. I mean, first of all, there’s a personal integrity thing.
Scott Wallsten: Well, I mean, but you could disclose exactly what it is.
Jason Furman: Yeah, that’s true. If I did, I think it would be that way. I don’t know that I even want to work hard enough to understand all 95 pages. I mean, there’s parts of it that use statistical techniques I’ve never used before, and I’ll bet they’re right, but I’m not positive about that. And I think that does speak to—I’m not sure anything it’s written is a brand new addition to knowledge that I am denying the world a journal.
If I was a cancer researcher, and it discovered a way to cure cancer, and I was like, but it has Claude’s name on it, and I’m not that familiar with the statistical technique, and I’d feel sort of bad, so this cure for cancer that I personally discovered on my computer, I’m just going to keep it on my computer, that would be sort of a bad attitude. None of these papers feel to me like they’ve made an earth-shattering discovery.
But that also gets to what they’re replacing. So I think of the papers my colleagues do at Harvard, and they are very novel, cutting-edge research. And I do think, for AI, that so far is harder, maybe not in theory papers, but for empirical papers. But then I think of what an awful lot of think tank work is, and it’s a little bit less about novelty and a little bit more about applying something to a current situation. If on the Hill they’re debating whether to raise the minimum wage to $15 or $20 an hour, you don’t want to write a brand new paper to answer that of original empirical research. You want to take a lot of existing elasticities and studies and whatever else and combine them, and the AI might be better at that second one.
A decent amount of my work, policy work, writing for the public, et cetera, is, at this point, definitely supported and informed by AI as an input to me that is different from a groundbreaking new discovery, which I haven’t seen in economics. But they have them in math right now. I would not assume that we’re not going to have them in economics.
Scott Wallsten: Jason isn’t putting his name on these papers. Someone with a client and a similar system might, so I asked him about the person on the receiving end.
So let’s say that we’re in the near future, where almost anybody on any side of an issue can produce a pretty solid study overnight. A lobbyist hands a congressional staffer a 60-page study on Tuesday, and markup is on Thursday. What should that staffer check?
Jason Furman: I don’t know the exact answer to that question. I mean, first of all, there’s even a question in the future of why the staffer would be interested in a study from anyone else at all. So if you’re in the think tank business, part of what you do is inform people who don’t have access to the analysis themselves. But that congressional staffer could have the 60-page document written for themselves in exactly the format, tone, length, coverage, etc., that they want.
But suppose they have handed it to them. Now the congressional staffer can use AI to help try to figure out the problems with it in a way that they couldn’t have before.
I’ve always thought that’s the most important quality when I was in government, was being able to ask people tough, skeptical questions. Now, I thought listening to industry is incredibly important because they know so much that you don’t know in the White House, but they’re also so self-interested. So you need to be smart when talking to them.
Scott Wallsten: You just raised whether think tanks or industry studies have any role at all. If anybody can do this, what becomes the scarce resource?
Jason Furman: I don’t know. I mean, look, I think there’s still a certain amount of judgment and taste. What’s an interesting question? What’s not an interesting question? Now, sometimes the question’s out there. So there is a question of, should you raise the minimum wage? Should it be $15 an hour or $20 an hour? Whatever it is. That’s not that interesting and complicated a question to come up with. The political system came up with that, but I think a lot of questions require a certain amount of judgment and taste.
I think credibility will be a scarce resource, which is if you put the name Claude on something versus you put the name Jason Furman or Scott Wallsten, I’d like to think some people are a little bit more interested in reading it if we’ve put our names on it. And maybe they know I put my credibility on the line and it comes out wrong, and I’m the one who’s going to look bad. But I’m also open to there being a point, and maybe not in that distant future, where, when it comes to sort of analytics that aren’t totally, totally cutting edge, there just isn’t a huge amount of edge that people have.
Scott Wallsten: And since you have the second system, Briefwright, tell us a little bit about that and how it differs, because it also produces some pretty long output, even though they’re briefs, right?
Jason Furman: So Paperwright was an experiment just to see what it’s like to write an economics paper from scratch. And my thought is that I’m going to just watch that over time. I’m going to see where it is six months from now, where it is a year from now. Briefwright is an actual active tool that I use for myself all the time.
What I like about Briefwright is it’s sort of what I’d ask an RA for, but might expect them to spend a few weeks on it, which is I want to look at something a lot of different ways. Now, in the wrong hands, I think this could be dangerous. I could look at something 50 different ways. One of them supports my argument, and I go write that one up and ignore the other 49. I’d like to think what I’m doing increasingly is a level of robustness I’d never done before, where, if 49 of them say the same thing, I write that up. And by the way, I might even post it so you can see the one that doesn’t support my view. And you can be the judge about the assumptions that went in.
Scott Wallsten: But the p-hacking question, which is now sort of entire paper p-hacking, becomes possible. Like you said, you’ve got the one paper that is different from the others. And it’s a great robustness check for people who are being honest about it. But a lot of people just want a particular result, and will choose that one. And maybe that’s the same problem we’ve always had. But again, it comes back to the same question of how will people know?
Jason Furman: Look, I mean, that gets back to the credibility as a scarce resource as part of how it knows, part of how we’ll know. There are certain economists who always know the answer before they do a paper. And there’s other economists who are so excited to figure stuff out that they go do papers when they don’t know the answer, because they want to learn what the answer is.
And it’s not like the former, if they knew the answer and write the paper, that I would discount it all the way to zero, nor is it the case, the person who’s curious, that it’s 100% sure that what they came up with is correct. But, as a Bayesian, I update more on the second type of person than I do the first type of person when they say something. So I think that will be one of the scarce resources.
The other thing is, my hope is that you can detect these problems more easily. I mean, you can ask an AI, hey, this person did it this way, what if you do it another way? And the AI can say, hey, I just did it 50 other ways, and every other way gives you the opposite answer to this one. And then it’s like, huh, maybe I don’t think that. So we’ll all be able to do more of it.
And then the final thing is, I do think a little bit less dead tree papers and a little bit more interactive living websites that grow and evolve and let people do more sensitivity analysis and more choose-your-own-adventure would be a good thing.
Scott Wallsten: What would you tell the agencies, like GAO and CBO and the Congressional Research Service? What should they be—should they be trying to reproduce studies? Should they be changing the way they evaluate programs? I mean, their reports are influential, traditionally, I think, very good. And they’ve kind of always been a sort of a quality measure on what we’re learning about different programs.
Jason Furman: So I actually have certainly shared my thoughts on this with CBO, including the director, Phil Swagel. And my biggest two things are, one, be doing a lot more experimentation. Two—actually, I’m going to have three. Two, be more ambitious. You can do bigger, more complicated things than you’ve done before. So ask yourself, all the ideas you had that were just too difficult and too tedious to do, can you pull them off now?
And then the third thing is, CBO—the one shortcoming it’s had is it’s not transparent enough, and it doesn’t let its users change assumptions as much as I think it should.
And one of the big arguments CBO directors have always made to me is, Jason, it’s actually really hard to do all the transparency stuff. It takes a while, and so we’re being asked to do so much stuff. Do you want us to do one quarter less, so that we have enough time to post the replication? And that’s much less of a trade-off now. I mean, I write, effectively, blog posts on Econvitals, and I have a huge replication package, which I just tell it to post. I don’t look at the replication package, but I assume it can replicate my blog post. So they can do more of that.
You can’t have every member of Congress choosing their own cost estimate. CBO has to have its cost estimate, everyone has to live by that rule. But CBO does analysis of inequality. No reason not to have a website, and you’re a Republican and want to see lifetime inequality, you’re a Democrat, want to see point-in-time inequality, you want to value Medicaid at willingness to pay, or at government cost, or at willingness to accept. Have at it. You should be able to do that analysis your own way. And this tool can both let CBO do all the analyses and let it figure out a better way to present it, so users aren’t just reading a report, but in some ways are writing their own report within a set of parameters and choices that CBO has, in its wisdom, set out.
Scott Wallsten: Coming back to your two systems. Do you get different kinds of results from Briefwright than from Paperwright?
Jason Furman: So Briefwright does more data description and less causal analysis. And so in that sense, it’s less likely to give you an answer and more food for thought. Why did this go up? Why did this go down? Paperwright is trying to write an actual NBER-style paper. And so it might have instrumental variables, or diff-in-diff, or all sorts of statistical techniques. And in some ways that makes me more nervous, because I just want to—are the things correlated? That’s an interesting thing to know. I understand when it’s correlated, it doesn’t mean it’s causation.
So I can then sort of, I don’t know, use judgment, use whatever else. Once it gets really complicated, I start to trust it less. But to some degree, that’s also true with papers. I mean, I am much happier with a paper that shows me the crosstabs, they tell a compelling story, then it goes through a whole complicated causal analysis, and it sort of confirms what you could see with your own eyes. Now, I’m not saying that’s the only thing in the world. Sometimes there’s all sorts of spurious correlations, et cetera.
But I like it when the simple and the complicated get you to the same place. And if I had to choose one, at least right now, from an AI, I actually would be a consumer of the simple, with all the caveats attached to it, rather than the complicated, which comes in some ways with fewer caveats and more confidence.
Scott Wallsten: What reaction have you been getting when you talk to people about this? Is there institutional pushback, or are people excited?
Jason Furman: I mean, the thing that got the biggest pushback was when I made the observation, which I think is sad but true, that I just think research assistants are going to be less needed in the future. That doesn’t mean no one should hire research assistants. That doesn’t mean they’re not needed. But it literally is the case I’ve not heard from a single person who says Claude Code is so powerful that my research assistants are now so amazing that I want to hire twice as many of them.
In my own production function, I am the scarce input. If you gave me—if I worked with people that gave me twice as much stuff, I couldn’t process it. And so if you have a technology that can make twice as much stuff, I need half as many people. So the research assistant part has been controversial. A lot of people have reacted against it.
Some of those reactions I agree with. Where does the pipeline come from? I wouldn’t be me today if I hadn’t been a research assistant, and maybe we need to figure out some answers to that.
Scott Wallsten: I’ve found the same thing you’re observing. I looked at research assistant credits on NBER papers, and they are starting to decline, especially in methods papers, suggesting that there are fewer RAs being used to produce those papers. It’s just suggestive now, it’s early. But I agree, I have not seen the “we’re doing more work, therefore we need more RAs” effect. And I also feel sad about it, because I don’t know where they’re going to come from. What are your thoughts on that? Where will the pipeline start now, and what should it even look like?
Jason Furman: Education obviously still plays a role. It’s not like people aren’t teaching their students and mentoring their students. That is literally my job, and I take that job very, very seriously.
And in some sense, that says someone’s paying for you to be educated. It might be you paying your tuition. It might be the government paying the tuition. It might be some other way, and that’s a totally viable way to do training. And to the degree training is a private good that somebody themselves is going to eventually benefit from, maybe they can pay for it. If we think there’s a public good aspect to all of this and a collective action problem, then maybe the government pays for some of it.
I think we can all have a thumb on the scale for mentorship and what we do, but you need to be a little bit hard-headed about it, too. I’m a foundation that’s all about curing cancer, and I’m giving grants to cancer researchers. And somebody can do more clinical trials or have more RAs. And they tell me if they do more clinical trials, they will cure cancer faster, but if they have more RAs, they’ll be better cancer researchers 25 years from now. I think most people want the first, not the second.
Plus, we don’t know how many researchers we need 25 years from now. And that’s the last painful thing in all of this, is there’s less of a pipeline in, there might be less of a need on the other end as well. And certainly, I don’t think anyone should be paid just to bot-sit and have AI write literature reviews for somebody else and pass it along. It has to be real and meaningful work. Otherwise, what’s the point?
Scott Wallsten: Are you finding that the students now, graduate students, are kind of differently focused than they used to be? I mean, one outcome I could imagine is that younger people to the field learn new ways of using AI that we would never even have thought of, because we weren’t natives. Are you seeing that at all, or is that just wishful thinking?
Jason Furman: If you’re talking about our PhD students, there’s enormous variance in them. Some of them are all in, hugely in. And others are sort of more chatbot, maybe helping it with cleaning the data and stuff like that. But I think they all understand that to get a good job three years from now, they’re going to need a much better, much more ambitious paper than you would have needed to get a job three years ago.
And the types of one-shot public data papers that my system writes, I mean, those couldn’t have gotten a good job five years ago, but they could have gotten a good job 25 years ago. They would have probably been impressive by the standards of then, but two years from now, people are going to expect much, much more.
There’s also something interesting in economics hiring, which is people who are making the decisions about who to hire as junior faculty aren’t reading the papers that carefully. They’re going to the job talks, and they are watching how the person handles really difficult questions in real time. Not totally positive that that’s the most efficient system. There might even be a little bit of laziness that goes into that, but that’s sort of the system.
And I could see more things educationally, academically, et cetera, going that way, that what you expect from people is not that they produce really polished, impressive writing, but that when you ask them about an assumption, how something will change if an assumption changes, why they made this choice, what the implication of that is, that they’ll have really interesting answers. And to go back to your Hill staffer at the beginning, maybe that’s what you’ll have to deliver to them, isn’t a 40-page paper, but a really interesting, engaging, persuasive one-hour Zoom call.
Scott Wallsten: Do we need to start thinking about whether there should be standards for AI disclosure, and what should go along with the paper? I mean, it took a long time for journals—and not all journals even do this—to require code and data to become available for any published paper. Do we need something like that for AI?
Jason Furman: I’m not sure. I am disclosing myself wherever I can. I do find other people’s disclosures a little bit tedious. I mean, I saw one paper recently that had a slide that had eight rows, and each row was zero to three about how much AI was involved. And it was coming up with the idea, downloading the data, running the regression. I was sort of like, I don’t really care. I want to sort of see the output.
But I do think that whatever we have has to be incentive compatible, and the incentive to use AI is really high. Mostly that’s because it makes things much, much better. It’s a good incentive, and it’s very much aligned with the output we want. In some cases, it’s because it makes it easier and lets you do cognitive offloading, but whatever it is, trying to have no AI policies for things where people are going to use AI is just asking everyone to be a liar and to be a cheater.
So I think, ultimately, we want the best possible work, and we want the person whose name is on it to stand behind it. And ideally they’ll have opportunities to present it orally, and hopefully they’ll understand what their AI wrote for them.
Scott Wallsten: That’s a good place to leave it. We’re all figuring this out as we go. It’s scary, but it’s also fun.
Jason Furman: Absolutely. I feel like a combination of thrilled and terrified. And for myself, the thrill is higher than the terror. For younger people, honestly, the terror is a little bit higher than the thrill.
Scott Wallsten: Understandably. All right, Jason, thank you so much. I really enjoy talking with you.
Jason Furman: Thank you.
Scott Wallsten: A postscript. Jason and I talked about a paper his system wrote the morning we recorded. I’ve since read it, and it’s a great illustration of what we discussed.
Jason typed a three-line request asking what the UK’s large minimum wage increase did to jobs, income, and poverty. Three hours later, he had a 90-page paper.
On jobs, the paper found two answers that disagree, and it said explicitly that neither one measures the effect of the policy. It checked 63 outcomes to see whether the areas it compared were already on different paths before the policy. In 40 of them, they were. The paper concludes that its job results are descriptive and can’t be read as the policy’s effect.
The system wrote its analysis plan before it looked at results. It then amended the plan five times after seeing results, and the paper lists each change. By its own account, its headline numbers are exploratory. The memo that came with the paper adds that in the second round of AI review, two referees rejected it, and three asked for major revisions. They put its chances at a top journal at about 2% before the final revision.
So here’s one answer to the question I started with. The paper reports where it was weak because Jason built his system to be honest and straightforward about that. The main caveats are in its abstract, but someone else might not build their system that way.
Scott Wallsten is President and Senior Fellow at the Technology Policy Institute, a Policy Fellow at the Stanford Institute for Economic Policy Research (SIEPR), and a senior fellow at the Georgetown Center for Business and Public Policy. He is an economist with expertise in industrial organization and public policy, and his research focuses on competition, regulation, telecommunications, the economics of digitization, and technology policy. He was the economics director for the FCC's National Broadband Plan and has been a lecturer in Stanford University's public policy program, director of communications policy studies and senior fellow at the Progress & Freedom Foundation, a senior fellow at the AEI–Brookings Joint Center for Regulatory Studies, a resident scholar at the American Enterprise Institute, an economist at the World Bank, and a staff economist at the U.S. President's Council of Economic Advisers. He holds a PhD in economics from Stanford University.
Jason Furman is the Aetna Professor of the Practice of Economic Policy, jointly at Harvard Kennedy School and Harvard's Department of Economics. He serves as the Weil Director of the Mossavar-Rahmani Center for Business and Government at Harvard Kennedy School. He is also a nonresident senior fellow at the Peterson Institute for International Economics. He co-teaches Ec10, "Principles of Economics," the largest course at Harvard. From 2013 to 2017, Furman served as the 28th Chairman of the Council of Economic Advisers, acting as President Obama's chief economist and a member of the cabinet, capping eight years as a top economic adviser in the Obama White House. His research and writing span U.S. and international macroeconomics, fiscal policy, labor markets, and competition policy. He is a contributing Opinion writer at the New York Times and the editor of two books on economic policy. Furman holds a Ph.D. in economics from Harvard University.