AI Can Produce Research. Who Will Produce Researchers? (text only)
I’m helping destroy the ladder I once climbed.
I learned empirical research by wrestling with raw data, making mistakes, and having more experienced economists explain what went wrong. Later, I hired research assistants who learned the same way.
This fall my newest assistants helped me finish a hundred-page study of rural broadband subsidies in two weeks. They assembled data, wrote code, proposed statistical models, and drafted text. A second set independently re-estimated the results to check them. They cost me a few hundred dollars.
They will never get Ph.D.s.
A project like that would once have taken me a year and several human assistants. The assistants helped produce the research and learned how to do it along the way. AI made the research cheaper. On this project, it made the training disappear.
Last year I warned that AI was breaking the bottom rung of research careers. The risk extends to any field where beginners learn by doing work AI can now do, from biology labs to law firms. I now wonder how the next generation will acquire expertise when researchers like me can increasingly do without their help.
Some research tasks may now require less financial support. But concluding that we can stop investing in researchers confuses the cost of using expertise with the cost of creating it. A subscription buys access to the models, but not the years of experience it takes to direct them well.
AI makes producing an answer cheaper, so we’re going to get more of them. Two finance professors recently generated 380 manuscripts in about 12 hours with AI. Governments use research to decide whether programs such as broadband subsidies work. Producing an estimate is only part of that job. Someone must recognize when a persuasive result rests on a misleading comparison.
Yet AI could also improve the apprenticeship. I nearly missed that possibility this summer.
A colleague and I had asked a postdoc fresh out of graduate school to update a paper we had coauthored in 2010 on how companies buy and lease airwave licenses from one another. The update would provide information useful to policymakers while giving him a publication and a new industry to learn.
A few weeks later, stuck on the floor of a crowded Denver terminal waiting for a delayed flight, I had an AI model redo the whole paper. It took about 30 minutes. Without AI, I worried, we were entering the postdoc in a stock-car race with a horse and buggy.
My coauthor knew better. He asked the postdoc to check the AI’s work. A second model could probably have done some of that, but by doing it himself he came to understand the data better than either of us.
At our next meeting, the postdoc brought a dozen charts and said we needed to cut them to four. We spent an hour working through which ones. He brought the analysis and a clear sense of what it needed. We brought years of watching the industry. Together, we considered whether a surprising trend reflected a lapse in the regulator’s authority to auction new spectrum licenses, and which people in industry and government could help explain what we were seeing.
Neither of us had touched the project since the previous meeting. He had done the work. Our hour was the apprenticeship.
The catch is that my first instinct in Denver is the one the new incentives reward.
The old arrangement gave senior researchers a practical reason to teach. When an assistant made a mistake, I explained it partly so they would learn and partly because I needed better work next time. When a model makes a mistake, I fix the prompt and finish the paper. Even if someone else pays a student’s salary, training requires my time. A junior professor facing a tenure clock has a powerful incentive to skip the human and choose the route that publishes faster.
Academic departments, funders, and firms can no longer assume that training will emerge automatically from senior researchers’ staffing decisions. Nobody yet knows how best to train researchers alongside AI. We do know that experienced researchers’ time is now the scarce input, and that the people who learned the old way will not be around forever. Departments and firms will need to reward those who spend that time training others, because so many projects no longer include it.
We also need to measure what actually works as research roles evolve. That makes the administration’s fiscal 2027 proposal to eliminate the National Science Foundation’s social, behavioral, and economic sciences directorate particularly ill-timed. The directorate provides much of the federal money for academic social science, which pays many research assistants. It also houses the science of science program, which studies the scientific workforce. We risk weakening the apprenticeship while cutting our capacity to discover what should replace it.
I have no desire to return to the yearlong project. I want the faster analysis, the inexpensive checks, and the ability to pursue questions I could not afford before.
My AI assistants may never become my colleagues. Someone still has to train the people who will.
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.
