AI Could Do My First Job. Not My Last.
Warning: This is different from what my other Occasional Posts will be. It is longer, more personal and more opinionated, with almost no data and no charts. But it also addresses some important issues, so I hope you read it. Or if you’re busy, paste it into an AI and have it produce a summary and critique at whatever length or depth works for you.
I want to talk about my comment to Neil Irwin about the White House Council of Economic Advisers (CEA) when Econvitals was released: “If I were running the CEA today and had a choice between 20 staff members who had a combination of current Ph.D.s and future Ph.D.s versus the AI systems that we have today, I would somewhat regretfully and sadly choose the AI systems.” It would be fair to say the reaction to it was generally negative, including from many friends and former colleagues. I confess that makes me a bit mad and a bit sad.
The mad is entirely at myself for choosing a provocative example in order to stimulate a conversation, when instead it backfired and closed off the conversation. And the sad is because we are facing some very difficult and challenging issues, at least in certain pockets of the labor market, and I worry that unrealistic denials of risks and tradeoffs—albeit not certainties—could get in the way of honestly confronting what we are facing. At a moment like this we need realism, honesty and a willingness to make difficult choices.
I am not saying my view is definitely right, just that it is not so obviously wrong. In this post I will try to do a better job explaining why and what it might mean. I am going to be granular about CEA because I think it is a useful thought experiment, in the ways it does and does not generalize to other organizations.
The Council of Economic Advisers: An Amazing Group of Thirty-Five Humans That Could Be Mostly Replicated by About Ten Humans Plus AI
So let me get to my specific claim: I believe that most of what CEA did when I ran it could now be done by about ten people. Which means AI could replace much of what a group of bright people, from recent college graduates to seasoned academics, could do. So my 20-person comment to Neil may have understated the case. (Later on I will get to the different and more important question: not reproducing what we did but figuring out what to do in the future, which shifts the issue from whether AI could replace people to whether it should.)
A bit of background. CEA is like an in-house think tank that provides economic advice to the President and the administration, as well as a set of public reports including the Economic Report of the President. It is a small, scrappy staff of about 35 people, most of them not permanent government employees—including about ten Senior Economists, many of whom are professors on leave from universities, and another roughly ten incredibly talented and hard-working junior staff who are often headed to or in the middle of Ph.D. programs.
When I ran the Council of Economic Advisers from mid-2013 to the beginning of 2017 we produced dozens of reports that helped contribute to the conversation on everything from competition to male employment to how to extract the signal from macroeconomic data, not to mention contributing to myriad policy issues including Medicare payments, criminal justice reform and the Clean Power Plan. None of that would have been possible without the ideas, initiative and execution of the people who worked there.
AI on its own or AI with just me directing it could not replace all of that. As I told Neil, “You would need people at CEA to interact in meetings, to talk to the press, talk to Congress.” But CEA would not need nearly as many humans for just these tasks.
AI is an outstanding research assistant and is doing increasingly sophisticated economic reasoning and analysis. But humans would still play a critical role. If I were running an AI-intensive CEA with only a skeleton staff, I would expect the remaining humans to be in frequent conversation with economists and policy wonks, both across the administration and outside it, to challenge our ideas, get new ones, share our results and try to persuade people. This is not complete automation. But it is AI doing most implementation.
The Messy Reality of Who AI Can and Cannot Replace
My expertise on these issues comes only partly from my time as Chair under President Obama. Perhaps even more relevant is what I learned and did as a CEA staffer under President Clinton. At that earlier time my job was mostly performing short analyses, tracking macroeconomic data and writing short memos on it, drafting reports, working on the administration’s forecast, and working with Joe Stiglitz and then Janet Yellen on their speeches. It was amazing and challenging work. I would not be the person I am today if I had not gone through that intensive experience and the mentorship I got from them. But I look back on that time and truly believe that AI—guided and directed by Stiglitz, Yellen and other more senior people—could easily do 95 percent of what I did then. And the other 5 percent was staffing Stiglitz on trips, doing things like getting him a new tie when he spilled on the old one.
A few people responded to my argument with claims that the CEA Chair might even be more replaceable than many of the people who work for them. It might sound just, but it is also wrong. Jason as Chair could not be replaced by AI but Jason as junior staffer could be. That is not because Jason as Chair was smarter or better than Jason as junior staffer. It is because the task mix for the Chair is mostly about being in meetings, briefing the President, testifying to Congress, making speeches, reaching out to discuss issues with academics and the like—those just are not jobs that an AI can do. In contrast, the task mix for a junior staffer—and even for some senior economists with decades of experience—is mostly doing analysis and writing on a computer and emailing it to the people who use it. That could be done by AI.
The reality gets very messy in ways that measures of AI exposure are unlikely to capture. For example, the effect of AI on employment at the National Economic Council (NEC), where I did two stints, could be quite different from its impact on employment at CEA. The NEC and CEA are both White House economic advisers and employ similar types of people, especially in the junior ranks. But the NEC is a more strategic and political organization that manages the policy process and is heavily involved in the passage and implementation of policy. Most junior staffers at the NEC are probably safer from AI than even more senior staffers at CEA, not because NEC staffers are smarter and more talented than CEA staffers but because much of the job of NEC staffers is running meetings, building coalitions and working with a wide range of units to implement policies—none of which can be done by AI.
So the issue of who AI can replace is not one of who is more virtuous or smarter or better or even who is more senior or more junior. It is about what the nature of the job is. AI poses more of a threat to some of the smartest people in the academy, like mathematicians, than to people in many other disciplines, although I have heard a number of colleagues in different disciplines wonder if their days as active researchers are numbered.
Don’t Underestimate How Capable AI Is: It Can Write Decent Economics Papers, Very Good Briefings and Excellent Blogs With Basically No Human Input
Part of the disconnect in the discourse is that people have different understandings of the current capabilities of AI. More and more people understand them, but I still sometimes hear people talking about AI hallucinations or how it repackages Wikipedia. These might have been true of badly used, weaker AI models a year or two ago, but they are barely relevant for understanding agentic AI today.
To understand it better, as an experiment I built an engine I call Paperwright. It takes a short prompt and spends two to ten hours turning it into an economics paper, drawing on a multi-step agent process that includes researchers, writers, fact checkers and referees. For example, I entered “data centers and electricity prices” and about six hours later I got this 65-page paper, which has a mathematical model, an analysis of county-level data drawn from Environmental Protection Agency permit records, the Energy Information Administration and other sources, references and appendices, plus a replication package. I don’t think the paper is a great contribution to knowledge, and it should not be published in even a mid-level journal, but if it were a senior thesis it would probably get a grade in the Summa range at Harvard. And if you think that paper is wild, check out this one, produced with a one-word prompt: “Triffin.”
I also created Blogwright as an experiment to see if AI could write think tank-style blogs that explain things, organize and summarize data, or make a case. The results, in my judgment, would be among the better versions of the genre at the better think tanks, on topics including the case for reviving the Premium Tax Credit, how work from home has changed the labor market for women, why the budget problem is that taxes are too low and why the budget problem is that spending is too high.
These blogs are better than what a chatbot could do today, given the ways I have architected the multi-step process for writing them and the links to over one hundred data sources. But I expect even a simple prompt into a chatbot will do better than this in a year or two. Moreover, what started as an experiment has turned into an input into my thinking—on many topics I would rather get up to speed with an AI-generated blog post than anything written by humans, which might be out of date, not match my interests or have inaccurate or misleading numbers (AI, used properly, can make fewer mistakes than people).
I also created Briefwright to write briefings for me on various topics, ranging from literature reviews on audit studies to analysis of the wealth effect in supporting current consumption to an evaluation of the climate forecasting track record. None of them are fully reliable from beginning to end, but all of them are informative and generative as an input to my own thinking and writing.
Econvitals itself is built the same way. Almost every email you receive from it was written by AI. I don’t approve, read or edit them before they go out. I iterated on the prompts, designed the pool of charts they draw on and architected the overall system. The exceptions are a paragraph I write in Morning Notes about a chart I select and design, and these Occasional Posts. The emails have been timely, reasonably comprehensive and generally accurate.
But I also expect my own posts will be more interesting than anything the AI writes. I don’t say that to brag about myself; most days the Morning Note links to two pieces it selects as interesting by different people and organizations, and those are also more interesting than anything written by AI. As are innumerable other human-written pieces it does not link to.
Some of the AI-generated papers, blogs and briefings are cutting-edge. Some need significant work and have big deficiencies that need to be remedied by humans. And some of the longer papers might even be garbage. But all of them were done by AI in 2026 (it will get better), by systems I architected (I’m decent at this but many people are much better), and with basically no human input or iteration beyond architecting the system and then entering a simple prompt.
I can think of people I know who overrely on AI and produce slop they don’t understand. But I can think of many more people—including some economists and think tankers—whose error goes in the opposite direction: underusing the technology in ways that lead them to make too many mistakes and not think ambitiously enough about what they could accomplish. And if you’re only making errors in one direction (underreliance or overreliance) you’re using the technology wrong.
AI Can Also Enable Us to Do More and Better: But How Much More and Better Will People Want?
CEA was swamped with just 35 staff covering every single economic issue. If I were in charge I would not have ten humans plus AI replacing the CEA I knew; I would have something between ten and 35 humans plus AI doing something much better than CEA did when I was there.
One factor that influences whether technological change increases or decreases the demand for labor is how much demand there is for the output of that labor (more technically, how elastic that demand is—so how much does it increase when the price falls). A far smaller share of American workers are farmers today than two hundred years ago because of the combination of technological progress in farming and the fact that people don’t eat much more than they did two hundred years ago. In contrast, businesses use vastly more accounting services, so that even as accountants enjoyed huge productivity improvements we ended up hiring more and more of them (although that has fallen off lately). The latter effect is known as the “Jevons paradox”—when making something more efficient and cheaper leads to a huge increase in demand for it, so you end up using more.
In the case of CEA, I am not sure how elastic the demand for its output is—after all, it was already free. We would not have written more reports or memos; we could have covered a few more issues and done better analysis, but there is a limit to that as well. Plus many of our clients—in the NEC (which is less technical), the Hill, the media, etc.—could do a lot of this without CEA, so it is possible that demand for CEA’s output could go down, not up.
That said, in most domains we have consistently underestimated the amount of demand for new and improved versions over time. We can usually picture a living standard slightly better than ours, maybe twice as good, but have a hard time picturing one ten or a hundred times as good. And yet that is what we end up wanting and producing—which is why so many people still work so hard even though they could easily have an amazing living standard by the standards of 1930 with many fewer hours of work per week. So it is possible I am underestimating just how much more sophisticated the analysis will get—and how it might even take more people to work with the AI on that analysis.
But I want to be honest and realistic here, even though it’s hard. I have talked to dozens of colleagues at Harvard and other institutions. I have not met a single one who cites the Jevons paradox in their own hiring of research assistants. No one who says, “Now that research assistants have access to Claude Code they are so effective that I want to hire more of them.” To be clear, there may be some I have never talked to who think this, but they are greatly outweighed by people who think they need the same number, fewer or perhaps even none. I personally no longer have a full-time research assistant but continue to work with a few outstanding part-time research assistants who contribute a lot of ideas as well as execution to what I do.
Going beyond anecdata, the evidence itself is mixed. One paper by two of our graduate students at Harvard, Seyed Mahdi Hosseini Maasoum and Guy Lichtinger, found that junior employment is down about 9 percent at firms adopting AI. But Martha Gimbel and the team at The Budget Lab at Yale do not see the occupational mix changing in ways that clearly line up with AI. It’s still early days.
I would not be surprised if the optimum was a CEA about half its current size, staffed with more senior people, selected to a greater degree than in the past on their ability to operate interpersonally (so in meetings with other administration staff) and their ability to manage AI agents. And as AI gets more capable and the government updates how it operates, the optimal number would be more likely to go down than up.
(I also think that people who like and believe in government being capable of doing good and important things have a responsibility to help make it as efficient as possible. That will not always mean cutting staff relative to what we have now. Statistical agencies, for example, are vastly underresourced and nothing on Econvitals would work without the thousands of people who work at them collecting and analyzing data. That said, where we can do more with less—and there will be many places that is the case—we should.)
The Difficult Question of How Much to Give Up to Cultivate the Next Generation of Talent
So far I’ve talked about the work that CEA does in producing analysis and public policy. But what about the work it does in helping to train people who have gone on to everything from winning the Nobel Prize and the John Bates Clark Medal to serving as Treasury Secretary and Director of the Congressional Budget Office? You can debate the causal impact of CEA in each of these cases—much of it is selection, not treatment. But the causal impact is not zero.
We need to take the pipeline seriously and be mentoring, training and investing in the future. The best paper I have read on this topic is “Pyramids, Diamonds, and Oscillations” by Guido Friebel and coauthors, a model in which AI makes firms freeze junior hiring in the short run even if they eventually return to the familiar pyramid of many juniors and few seniors, while faster learning could shift them for good toward a diamond, with fewer juniors and a bulge of experienced people in the middle. I highly recommend it—or if you are not super technical about equations, I recommend you give it to your AI to rewrite in whatever format you most want to read it in (which could even be a music video).
But we also need to be honest about the tradeoffs. CEA was paid for by taxpayers. I could maybe justify hiring someone at $75,000 who produced $50,000 a year of output on the grounds that it would help train and mentor people to contribute more in the future. But I could not justify jobs where the person made only small contributions here and there and the bulk of what they did could be done by an AI, just because it would be good for them and the world more broadly if they got that training. Moreover, it is not clear what it means to get “training” if all you end up doing is botsitting.
Similarly, if a researcher gets a grant for economics research or cancer research, the grantors are usually looking for the maximum possible research output—more discoveries and cures. Although there are programs aimed at supporting the pipeline and requiring dedicated training and mentoring, most are not looking for Principal Investigators to run projects that either are more expensive or end up doing less in order to devote additional resources to mentoring and training so that ten or twenty or thirty years from now economics research or cancer research is better.
In universities our job is mentorship and training people for the future. I take that seriously, as do most of my colleagues. I also think organizations around the country will need to put some weight on that too. But there is a limit to how much weight.
How it works out in the workplace will be much less clear. Gary Becker explained that while firms would not pay for training in skills when workers could take them elsewhere, the training will still happen because the workers themselves will pay in the form of lower wages. That works fine if an investment banking associate is producing $300,000 of value for the firm but is willing to work for a mere $200,000—while de facto giving up $100,000 in exchange for the valuable training they are getting. But if the investment banking associate is producing effectively zero for the firm but getting valuable training they will take to another firm in the future, then do we really think it is a viable equilibrium for them to pay the firm $100,000 a year for the privilege of being trained?
The government is not profit-maximizing and can care some about the social good it does. But it is hard for me to believe that a make-work program for junior economists who are not actually producing much of value (because the AI is producing it) would have large enough social returns over time to justify its upfront cost. Moreover, if the government has spare funds to hire extra people it does not especially need in the short run, then economists should do their objective best to figure out where those skills are needed (public health? AI safety? National security?), not just assume that their own profession should be the one to be protected using extra taxpayer dollars.
I don’t know the full answer to the pipeline question. I worry about it. But I don’t think anyone should think it is easy or obvious or that no one needs to worry. At a minimum, we will probably need to be intentional about paying for the pipeline explicitly, through fellowships and training grants, instead of hoping that organizations keep hiring people they no longer need. But we may simply need fewer of certain types of skills and activities in a future world where AI is becoming increasingly capable.
I Don’t Know the Right Ratio of Joy and Terror, but You Should Be Feeling Both When It Comes to AI
I have generally been an AI accelerationist. When I was the Chair of CEA a common refrain in my remarks on the topic was this: “The biggest worry I have about AI is that we will not have enough of it.” My feelings now are much more complex, and if there were a button I could press to undo the last year or two of AI advances—and freeze the technology there for the next several decades—I would be seriously tempted to press it.
I personally feel alternately exhilarated at all the things AI can do for me and humbled and slightly depressed at everything I used to pride myself on that it is vastly better at. I feel excited for students using AI to make new advances in research and thinking but also depressed that they will not have to struggle, learn and grow in the ways I did, without which I could not imagine being who I am today.
I look at younger people setting out to do knowledge work and feel both sentiments, but with a higher weight on the scared and depressed, fearing that they won’t need to struggle and learn and that career paths and possibly even entire careers will no longer exist. Manufacturing workers and communities have dealt with much worse for decades, but it would not be truthful to deny feeling it a bit more strongly when it comes to one’s own family, friends and community.
I said I would be tempted to press the button, but I would not actually press it (although slowing a bit to get safety right would be a good idea). In part because there is no button. But mostly because the people I’m most worried about are at the upper end of the global population in terms of income or the meaning and purpose they get from their jobs. They have a lot of downside from this technology. But people who are much poorer, who suffer from terrible diseases or who will be adversely affected by climate change all have more upside than downside from more innovation, technology and productivity growth.
Regardless of our feelings about the button, the best advice for any person, organization, country or the world is to figure out how to use AI better while also thinking about what you can uniquely do that is not easily replaced by AI. Continuing to challenge yourself so your skills are not offloaded to AIs, while also building the full capabilities of AI into your life and work and improving your management and interpersonal skills and your unique, human value added, is the best shot you have for the future. It is not entirely clear how anyone can do all of this, and it may not be good enough, but I can’t think of anything better.
As for the broader public policy response, well, this post is long enough already—I’ll come back to that another time, likely in another venue.
Jason’s Occasional Posts are written by Jason himself; almost everything else on Econvitals is written by Claude. Subscribe to get them by email, or read the others.