What this covers
Tyler Cowen and Russ Roberts discuss AI's economic and educational future in this EconTalk conversation. Cowen argues that mass unemployment from AI is implausible, grounding his case in Say's Law—the principle that increased production of goods and services generates the incomes necessary to purchase them. He contends that even as some jobs disappear, new ones will emerge from energy demands, regulatory oversight, and entirely unforeseen categories, much as the Industrial Revolution created roles no one could have predicted. The conversation centers on what higher education should do in response: Cowen advocates that colleges devote roughly a third of course-time to teaching AI use, since nearly every future job will require it. He presents a practical solution to academic cheating through in-room, sampling-based testing, arguing the problem is overstated.
Beyond employment economics, the discussion ranges across how writing and thinking will change in an academic context, whether the value of college lies in subject mastery or in helping students become who they wish to be, and how individual adaptation to AI tools is already underway. Roberts raises skepticism about whether AI can replicate the internalized wisdom that comes from small-group struggle with difficult texts. Cowen identifies his deepest concern not as economic collapse but as political failure—worse governance, regulatory incompetence, and the political power of displaced white-collar professionals to block AI's integration. He also addresses the likely obsolescence of most educational technology wrappers around foundation models, and sketches a radically different educational equilibrium where AI certification displaces traditional credentialing. The conversation assumes familiarity with neither advanced economics nor technical AI literacy.
Cowen argues that AI will not cause mass unemployment but will create new jobs and wealth, and that higher education should radically embrace AI as both a subject to teach and a teaching tool, with the main risk being political/governance failure rather than economic collapse.
- Say's Law logic: AI-driven production of goods and services generates the incomes to buy them, so mass unemployment is implausible
- Colleges should devote a third of course-time to using AI well, since nearly every future job requires it
- AI cheating is solvable by sampling-based in-room testing, removing the main objection to AI in education
Education should embrace AI through market competition, flexible timing, and direct student-AI interaction rather than rigid institutional bundles.
- Following Adam Smith's principle that classes and professors should compete, colleges should let each student take one AI-driven class (perhaps once a year or less), possibly in a small self-recruited group on a topic the school doesn't otherwise offer like Tudor England, and let a market discover whether students want more or less AI—if they don't like it, they'll stop.
“Each year, let a student take one class with artificial intelligence. No more than one... and see what they think.”
- An AI-driven class can replicate a human course by working with a non-expert coach: the AI generates a reading list, quizzes, and paper assignments and grades them, optionally with a human re-grading to reassure people, all at zero marginal cost to the student—and if it lacks human warmth, insight, or depth that matters, students simply won't take it.
“you prompt the AI; it generates a reading list. You go off and you do those readings. You prompt the AI to generate quizzes... The AI can grade the quizzes for you.”
- Colleges should devote roughly a third of total course-time to teaching students how to use AI well, because almost every future job will require this scarce skill, and doing so does not displace other content but enhances it—you can teach students to use AI to better understand Homer's Odyssey while teaching the Odyssey itself.
“almost every job in the future will involve knowing how to use AI well. In most schools, that isn't taught at all in any formal sense.”
- The 15-week class, four-year degree, and eight-semester structure are highly artificial; AI offers far more flexibility—a class on the Iliad could be read in six weeks before moving on—and we will move away from rigid time-batched learning, which contradicts how experts like Cowen and Roberts actually learn through life by picking up and stopping topics as they wish.
“we are going to see the whole 15-week thing as highly artificial”
- A more radical and better future model is for a student to simply chat with one AI for fifteen weeks, then have a different AI grade that chat to assess what the person learned and assign a grade—eventually this is how AI education will work, though copying a human class is the safer starting point.
“you just chat with the AI for, say, three months, 15 weeks, whatever. And, at the end of it all, you have a different AI grade your chat with the first AI.”
- AI can easily outcompete universities at certification: have a person spend a day with an AI prearranged to test them across many areas, and you get far better certification of strengths, weaknesses, temperament, and knowledge than letter grades; this is the future equilibrium, blocked only by institutions' reluctance to unravel the educational bundle.
“The AI can outcompete us in certification, easily.”
AI will create net job growth through energy demand, new drugs and devices, legal work, and unpredictable categories we cannot yet imagine.
- AI will produce plenty of new jobs because its expansion requires far more energy (which needs people), generates new ideas for drugs and devices that must be tested and pass regulatory review, and counterintuitively will increase demand for lawyers since many new laws must be written for AIs—work humans will insist on doing with AI assistance rather than letting AIs do it themselves.
“There will be plenty of new jobs under AI. Just look at the energy sector. To the extent AI takes off, we will need much, much more energy.”
- It is very hard to predict where future jobs come from, just as people at the start of the Industrial Revolution losing agricultural jobs could never have imagined that some of us would become podcasters; we are in the same position regarding AI's new job categories.
“If you go back to the early days of the Industrial Revolution and you tell people all these agricultural jobs are going away, would you have two people sitting around the campfire saying, 'Oh, yes, a lot more of us will become podcasters'?”
- Through normal economic mechanisms AI will produce more wealth and not fewer jobs; people at the bottom will get many services free or near-free, there will be more billionaires and mega-billionaires as small teams build high-revenue companies, and those new companies create new projects and many new jobs, so mass unemployment is not coming.
“in terms of normal economic mechanisms, I expect we will have more wealth. We will not have fewer jobs.”
- Once AI produces more goods and services, it is hard to reach an outcome where most people are worse off: the goods are sold, prices fall if needed, and the production, marketing, and distribution generate incomes that let people buy those same goods—a version of Say's Law—so Keynes's liquidity-trap hoarding scenario is not plausible in an AI-enriched world full of fascinating new things to buy.
“in terms of normal economic mechanisms, I expect we will have more wealth. We will not have fewer jobs.”
Students must learn to ask better questions and use AI as thinking tool, combining AI output with human-directed analysis and solo practice.
- Teaching people to use AI well is largely teaching them to ask better questions—what historical, economic, and contextual questions to pose about a text like Homer's Odyssey—and this is an unending task that explains why not everyone becomes a podcaster; one can even learn to ask better questions by using the AI itself.
“other people need to learn how to ask questions better. That is why we are not all podcasters.”
- Writing classes should split the tasks: force students to write with AI and grade the joint product against a higher standard, while also teaching them to write on their own—mainly as a way to teach thinking, since most people won't need solo writing for its own sake but will need to think, and writing is a great path to thinking.
“a big portion of those writing classes, you force the students to write with AI... 'Well, the standard for a good paper is higher. Use AI. You have to use AI.'... And you grade the joint product.”
- Cowen still lectures in his History of Economic Thought class because there is something about the vividness of human face-to-face communication, and he deliberately had students learn the Ricardian model from AI first, then teach it himself, asking them to mentally compare the two—making the comparison itself a core lesson.
“I think there is something about the vividness of human face-to-face communication.”
- We already know some version of one-on-one tutoring works acceptably well, as in Britain's enduring tutoring system, and it could be 10 to 20 times better with AI assistance; feeding a few years of data into the AIs would let them figure out how to improve teaching further, a possibility unavailable without AI.
“the British have a tutoring system to this day in many parts of the country... It works acceptably well. It could be 10 or 20 times better with AI assistance.”
Upper-middle-class professionals face real displacement that becomes a politically destabilizing problem for democratic governance.
- A serious worry is that the most displaced workers will be upper- and upper-middle-class white-collar professionals—such as consulting partners earning $1.4 million who might be reduced to $300,000 jobs at Houston energy companies—and because this politically influential group effectively runs the Democratic Party, their dissatisfaction creates a hard-to-survive political problem.
“it is possible the people who are most displaced will be the upper, upper-middle-class white-collar workers, and they will have to move to Houston and work for energy companies.”
- Consulting firms could deliver their human-relationship value with about a third of their current employees; in the short run there will be a consulting boom because everyone needs help integrating AI into workflows (even though consultants may not know how), but in the medium term demand for consulting services will decline.
“I think they could do that with, say, a third of their current employees.”
- Cowen's biggest worry, with or without AI, is that politics gets worse and that worse politics makes economies worse; adding AI to that mix creates many scenarios where the people in charge fail to regulate it well, because we have far worse political models than economic models and don't even know what the median voter wants on AI.
“AI will change governance in ways that are hard to predict. We have worse political models than economic models in general”
AI will increase leisure, which is mostly good; major disruptions are normal in transformative revolutions and may signal genuine progress.
- AI will lead to more leisure time, and there already is more because AI helps workers complete tasks more quickly in ways not reported to the boss; more leisure is mostly a good thing, though not for everyone.
“There is already more leisure time because tasks you do at work, the AI can help you with more quickly. It is just not reported to the boss that this is going on.”
- People on the whole dislike big change, and they did not love the Industrial Revolution at the time, yet it is arguably the best thing that ever happened to humans; AI will be like that—the more people are upset, the more we may know things are going well, even though those clued in will find the changes to jobs and their children's futures genuinely unsettling.
“Did people love the Industrial Revolution at the time? No. Some did. Is it arguably the best thing that ever happened to humans? Basically, yes. So, I think it will be like that.”
AI's governance value is real but useless if humans ignore it; near-term governance may worsen from spam and input overload.
- Leading AI models like Claude or GPT would on average give better governance answers than most current leaders or regulators, but the real problem is whether humans will listen; AI will be used only as an aid, and in the short run some governance may worsen as open input processes get overloaded by high-quality but pointless AI-generated comments and spam.
“I would vote for Claude or GPT [Generative Pre-trained Transformer] over most of our current leaders or even people in the regulatory apparatus.”
Young workers should develop skills in messy jobs with human contact and non-routinizable tasks that resist automation.
- Advice to today's 18-year-olds is to learn AI and seek what Luis Garicano calls 'messy jobs'—jobs involving many non-routinizable tasks, lots of face-to-face contact, and solving difficult problems with or caused by other humans—since these resist automation and carry a premium in the AI world.
“Tell them to learn AI. Tell them to look for what Luis Garicano called messy jobs, in a very good online essay.”
Cowen personally practices selective AI use for work while avoiding feeding sensitive data into systems due to bankruptcy and security risks.
- Cowen has adapted to AI competition—which competes with his writing more than his podcasting—by doing more podcasting and more personal appearances that AI cannot do, estimating he has reallocated at least half of his time usage already in response.
“I do much more podcasting because of competition from AI, which competes with my writing more than my podcasting. And, I do more personal appearances, which the AI can't do at all.”
- Cowen keeps his AI dialogues formal and scientific with nothing embarrassing and does not feed detailed personal information, emails, or confidential job, national-security data into AI systems—because despite a fairly high degree of trust, companies can go bankrupt, get hacked by China, or change, so caution is warranted for now.
“I would not put detailed information about your personal life into the AIs. Probably not.”
- Cowen now writes somewhat less but partly writes 'for the AIs'—treating them as his best, most patient readers who understand all background context—and deliberately builds a model of himself so AIs know what he wants and how he thinks, a model he and others can use in the future.
“I write for the AIs. I think, what do the AIs need to learn, and what do they need to learn about me? They are my best readers.”
AI cheating is solvable through simple incentive design without expulsion or campus surveillance.
- The AI cheating problem is much overrated and solvable purely by will: for 2-3% of a student's output over their college career, lock them in a room without AI and test them, and if in-room performance diverges dramatically from take-home work, investigate—this incentive against cheating works far better than current approaches without needing to expel anyone.
“The cheating problem with AI is much overrated. We are simply unwilling to do something about it.”
EdTech wrappers around base AI models will fail; people prefer direct access to foundation models.
- Most EdTech startups that modify base AI models with bells and whistles—delaying answers or walking through learning steps—will not succeed, because people will just use the basic foundation model they are used to; the equilibrium is one base model people work with directly, not specialized education wrappers.
“My intuition is none of those--few of those will succeed. The people are just going to use the basic foundation model.”
Truck driver displacement will be slow because jobs involve tasks beyond driving; universal driverless systems face higher marginal costs than expected.
- Driverless displacement of trucking will come relatively slowly because a trucking job involves much more than driving—loading and unloading cargo and handling points of contact—and although Tesla may displace Waymo as a cheaper driverless option within about ten years, universal systems often end up with higher marginal costs than expected once they must handle all possible problems.
“a trucking job has a lot more to it than just driving. There are all sorts of ways in which you load, unload cargo, and deal with points of contact. I think those changes will come relatively slowly.”
Small countries without innovation ecosystems cannot build indigenous AI; geopolitics will shift based on AI access and national security needs.
- For a country like Israel, AI carries a national security imperative—without it 'you are toast'—whereas a country like Brazil might be safe regardless; most small countries lack the option of an indigenous AI capability that Israel's innovation ecosystem provides, so geopolitics will change radically.
“for Israel in particular, there is a national security angle. If you do not have AI, I mean, you are toast. Now, if you are Brazil, you might be safe anyway. But you are not Brazil.”
Economic growth correlates strongly with human rights and dignity; AI wealth creation enables humanitarian progress beyond material abundance.
- Economic growth matters far beyond mere stuff because resources are needed to make lives better in humanitarian ways: very poor societies are less tolerant, grant fewer rights to women, and are worse on dignity and human rights, such that GDP per capita and non-GDP gains correlate by roughly 0.95—a strikingly high figure that is part of the core message of Stubborn Attachments.
“GDP [Gross Domestic Product] per capita and what you might call non-GDP gains, they seem to correlate by about 0.95, which, to me, is quite striking.”
AI will soon exceed most human interviewers; applicants now prepare using AI, forcing interviewers to generate novel questions constantly.
- AI will soon be better than most human interviewers and may already be, though probably not better than the very best human interviewers soon; because applicants now use AI to predict what an interviewer will ask, Cowen must constantly generate new questions, and AI's obsequiousness can be easily corrected by prompting it not to suck up.
“I think AI soon will be better than most human interviewers. It may well be already. I am not sure it will soon be better than the best human interviewers.”