My recent post on the utility of AI has led to a lot of folks messaging or commenting to ask me about my view of the impact of AI on jobs. For a number of reasons, I’m inclined to think that while AI will lead to a huge change in the kinds of jobs we have, it will not eliminate the majority let alone most jobs, even for those who work mostly with our minds and the usual tools—that is, the computers and printers that replaced the last set of tools these workers used: typewriters, ledger books, and legal pads.
Why Human Beings Will Remain Central to Higher-Level and Especially Creative Work
First, as my post pointed out, the usefulness of AI is very much dependent on the human questions we bring to it, which in turn rely on our knowledge of some field. AI may replace some lower-level work (with some potential problems I will return to at the end of this post). It may help us carry out analyses we think of doing. But because it mainly draws on the usual approach that is taken to dealing with the problems that are the core of most brain work, it will not replace the more advanced work that is the role of high-level professionals. And it is unlikely to make major leaps that lead to the invention of new ways of thinking about long-term problems.
This judgment is not just based on my own experience with AI. It is based, rather, on my knowledge of the work of creative people in the arts, philosophy, and the social sciences and also of scientific inventors and musical geniuses. Two things are found in those who do creative work in these fields. The first is passion. Creative people are driven to discover new things, try new approaches, and look at problems in a new way because, on the one hand, they care deeply about the answer and, on the other hand, they don’t take conventional approaches for granted. The second characteristic of creative people is that they are odd or, to be more polite, somewhat idiosyncratic. Creative people don’t want to fit in and don’t want to adhere to convention. Usually they have been raised by odd or different parents or have had teachers who are odd or different. Sometimes they are driven to overcome disabilities or insults to their ego that came about because of where they came from or the oddities they had developed earlier. Sometimes they are different because they come from minority cultures. Or they come from multicultural families that encourage them to create ideas that synthesize the approaches of these different cultures or, in the vernacular, create mash-ups of different ideas.
Whatever the explanation, truly creative people are different. But AI is based on a training in conventional approaches. That training rewards answers that are accurate by conventional standards. It is hard to see how an AI device trained in this way is likely to challenge conventional ideas in a truly creative way. It might be extremely helpful to a creative, unconventional human being. But it is hardly likely to replace that person.
AI Won’t Replace the Need for Human Judgment
It is especially unlikely to be creative in fields that are not cut and dried but require judgment and experience. It’s no surprise that AI is really good at chess and at solving sophisticated math problems. Those are the kinds of skills where success and failure are clear and thus where AI models can be trained by people with relatively little skill in the field themselves (or by other AI models). And even in these fields, whether AI is truly creative is questionable. AI has really surprised mathematicians by solving some yet unsolved problems that are widely regarded as important. It’s done so, it appears, not by making creative leaps to a solution but by being able to try a huge number of conventional approaches sequenced in a wide variety of ways. That’s impressive, and the ability of AI to do what would exhaust multitudes of mathematicians is a new and important intellectual tool. But what to my mind would be far more impressive would be if an AI mathematician created new, unsolved problems or conjectures that are widely regarded as important but are not yet solvable. That would be a far more important example of mathematical creativity than proving the conjectures of great mathematicians of the past. I don’t see how the brute force ability of AI will soon accomplish that.
Of course, most of the work human beings do is not all that creative. Still, as the examples in my previous post suggested, even normal everyday intelligent brain work depends on the kinds of insight that may not be available to most AI. So while AI is going to make high-level mind and brain workers more capable and efficient, I think it is unlikely for it to replace them.
AI Will Create New Jobs
AI itself will also create new jobs. Indeed, the best evidence we have is that AI has to this point created far more jobs than it has destroyed. The huge investment in compute—in chips, data centers, and power plants—has generated tens of thousands of jobs. And because of the relatively short productive life of the most advanced computer chips, they will have to be rebuilt again and again.
It’s not uncommon for new productivity-enhancing technology to create new jobs even as it destroys others. Think, for example, of the impact of electronic check-out systems in grocery and other stores. They have, no doubt, reduced the number of check-out clerks in American stores—perhaps by as many as 140,000. But they have created roughly tens of thousands of jobs for those who design, manufacture, program, install, repair, and supervise electronic check-out systems. These jobs almost all pay better and are more satisfying than working as a check-out clerk. And because they are higher paid, these jobs generate far more jobs for those who produce other goods and services than the check-out jobs they have displaced.
Both proponents and critics of AI have recently been talking about the moment in which we see “recursive self-improvement” in which AI models take over the training of new and more effective AI models. Some recent observers have called this view into question for a number of reasons. First, there appear to be diminishing returns to the productivity of AI training as more resources—more compute and more training runs—are employed. More and more training on the same conventional ideas that dominate computer-accessible information cannot lead to faster improvement in the abilities of AI. Second, training of AI models becomes as AI slop—the mistaken or even hallucinatory product of AI models—accumulates in the internet data on which AI is trained. And third, the most important advances in AI modeling have come from exactly the kind of leaps of creative imagination of the kind unavailable to AI for reasons I’ve mentioned above. (See, Michelle Kim, AI’s recursive self-improvement might not come so quickly after all, MIT Technology Review, August 18, 2026.)
on the the accumulation of AI slop comes For all these reasons, it seems increasingly likely that the moment of recursive self-improvement will never be reached. But even if we do reach it, the growth of AI will continue to generate many new jobs. Someone has to design and manufacture the computer chips AI runs on and build the data centers in which those chips are found. Someone has to decide where specialized training in one field or another is warranted and establish that training even if AI systems eventually take them over. (And for reasons I hinted at earlier and will explain another time, it is that specialized training that is essential to making AI truly useful. The notion that there is such a thing as “general intelligence,” whether natural or artificial, is rubbish.) Someone has to sell AI systems to businesses, develop methods for employees to use those systems, and train people to work within them. Even if AI takes over some of this work, there will undoubtedly be plenty for people to do.
Productivity-Enhancing Inventions Lead to the Development of New Goods and Services
The second reason to think that AI won’t replace most brain work is drawn from the history of previous productivity-enhancing inventions. The great inventions of the past didn’t just enable human beings to attain their previous goals and desires more efficiently. They expanded their goals and desires. The invention of the railroad and airplane made it possible to much more cheaply transport goods. But, even more importantly, it expanded the number of goods that it made sense to transport. When I was growing up, it was incredibly special to have fresh grapes, peaches, plums, nectarines, and cherries in the summer, largely because we only had each of these fruits for a few weeks. But now we have them most of the year because they can come much more quickly from distant places. And a whole bunch of fruits that, as far as we knew in the Northeast in the 1960s, did not exist, are now available to us much of the year, such as mangoes, guavas, starfruits, and many others. We coffee lovers no longer have a choice between Maxwell House and Chock Full of Nuts coffee but can try hundreds if not thousands of varietal coffee beans.
Much the same is likely to be true in brain work fields that are influenced by AI. We expect AI to help doctors become better and more efficient at diagnosing diseases and suggesting appropriate treatments. But it is also likely to vastly expand both the diseases that can be diagnosed early and the pharmaceutical and surgical treatments that can address them. Even more importantly, it is likely to attain one of the central aims of 21st-century medicine, to tailor treatments to our individual physiologies. AI may replace the work currently done by many doctors. But it will vastly expand the work that doctors can do. Even if 75% of medical work is done by AI, if the capacity of medicine to identify and treat disease early expands exponentially, we may well need the same number of doctors and other medical professionals that we have today. (And of course we have terrible shortages in many medical fields right now.)
And much the same will be true in other fields where AI will help businessmen, lawyers, investment advisers, and communications specialists not only work more efficiently but to do more and different kinds of work. When I was talking with an AI interlocutor about how AI could contribute to the work of a political organizer, I didn’t just focus on how we could do the stuff we usually do—such as producing media advisories, press releases, emails, and graphics—more efficiently. I focused on how much better these things could be if we could tailor them to the individual people we want to reach. That is, I thought about how much more and different work our comms staff could do, not just on how I could do what we do now with a smaller staff.
AI Will Lead to a Shift in Jobs to Those Fields Where AI Cannot Replace Human Workers
For these two reasons, I’m confident that AI will not replace most intelligent mental or brain work. But even if to some extent it does, I think there will still be plenty of jobs to go around. For it is likely that as jobs that AI can do well lead to the replacement of human labor, the added productivity will give us the wealth to buy goods and services that AI cannot do well. Again, the history of previous technologies shows that this is the likely result of AI. In 1860, on the eve of the Civil War, roughly 53 to 60% of Americans worked in agriculture. And this does not include those who produced, refined, and transported agricultural products. By 1900 the percentage of Americans employed in agriculture had shrunk to 36 to 41%. And by 2000 it had shrunk to 1.9%. (Another 10% were involved in processing, refining, and transporting agricultural products in 2002. It’s likely that this percentage was higher in 1900 and 1860.)
Yet despite this decline in agricultural employment we produce far more and varied agricultural goods than we ever did in the past. And we do so not just with fewer workers but much less land.
Despite this massive reduction in the number of agriculture workers, the majority of Americans work and our unemployment rate in the last ten years has been relatively low, although it has been higher for Black and brown workers. The reason is obvious. We consume a huge range of goods and services that did not exist in 1860 or 1900 or even in 2000. We can afford them because the huge productivity growth in agriculture as well as in manufacturing has made us much richer. So we can afford not only a wider range of agricultural products but far bigger houses, more and more powerful and comfortable cars, the latest entertainment and recreation technology, far more medical, psychological, and senior care, far more education, and so on.
So if AI reduces the number of workers in some fields, we can expect that we will need more workers in fields where AI is not likely to be effective. Are there such fields? Yes, of course. Personal and medical care is one example. It is easy to envisage a world with far more mental health therapists, social workers, doctors, nurses, physical and occupational therapists. Education, not just in our early lives but over a lifetime, is another. Human beings seek to learn and with time and enough money we will seek more teachers in a wide range of fields. The creative arts and recreation are two other areas where we can expect employment to grow. (And yes, I know that employment in the arts has been declining. For reasons I will explain another time, I believe this is a temporary phenomenon that has little to do with AI.) As we become richer we will seek out more, and more varied, forms of entertainment and recreation.
And I’ve only mentioned the expansion of jobs that already exist. There are undoubtedly goods and services that will come into existence but that I’m incapable of conceiving now. (If I were, I’d go invent them and make a bundle!)
We should not expect that this transformation will be automatic. Consumer spending on entertainment and recreation as a share of GDP doubled between 1958 and 1999. But it has remained flat since, largely, I suspect, because of growing inequality and rising spending on home entertainment technology and computers, which diverted spending from in-person entertainment. But it would be odd if spending in this area did not increase as we grew richer, especially if the efficiency benefits of AI are spread widely.
Now it is possible to imagine all kinds of ways in which AI could automate some of these fields. I would be surprised if AI did not find a major role in education, and medicine, and perhaps even in mental health. It may replace tour guides. And it is already replacing the work of some visual and musical artists. But again we have to imagine a world not just in which we consume the same goods and services that we do today but in which the goods and services we consume are more extensive, more varied, and more individualized. Perhaps AI will be able to do some intake work in mental health or help people with mild adjustment problems. It is likely to provide some low-level generic education and training. But as this work becomes more advanced, and as it provides more individualized services that require creative insight into the special problems and prospects of human beings or that seek to motivate human beings to attain deeper and higher goals, the human touch is likely to be profoundly important. Most of us can remember a few teachers whose entire character and being touched us deeply or whose creativity inspired us to fall in love with certain fields or approaches to pursuing knowledge. An artificial life form with the capacities and insights of a human being might be able to do that. A computer screen and disembodied voice is unlikely to do so.
AI Will Reduce the Work Week
And finally, even if my optimism about the growth of new goods and services and my pessimism about the ability of AI to give us the individual touch that human beings at their best provide is misplaced, the final reason not to worry about AI replacing most human jobs is that we use the great productivity enhancements of AI to do what we did in the 20th century and drastically reduce the work week. If AI is as effective as its acolytes think in increasing the efficiency with which we produce the current goods and services we produce, I doubt that many of us would complain if we made the same income by working only 20 or 30 hours a week.
Inequality, Equality, and AI
To this point, I’ve mostly avoided one critical subject, the impact of equality and inequality on how AI changes the world of work. The question of reducing the work week makes that question inescapable. Sharing fewer jobs among the entire population will no doubt require legislative action as it has in the past, mandating overtime pay for work done above the standard work week.
But that is not the only place where the question of equality comes to the fore in thinking about a future with AI.
There are many reasons that it is important that we focus on inequality when we think about the transition to a new AI-influenced economy. It’s fairly obvious that even under the most promising scenarios, the transition to that economy could permanently damage the lives of people whose jobs are displaced if they do not receive the training and temporary assistance they need to secure new jobs. Good tax policies that slow the transition and that raise funds for re-training programs are necessary here.
Another potential concern is that AI will displace entry-level jobs that serve as the lower rungs of career ladders. For, as I’ve suggested from time to time, AI is more likely to be effective in doing lower-level, more routine work, than higher-level, advanced, and more creative work.
This is not just a problem for young workers but is a problem that could terribly damage our economy as a whole. The experience gained in entry-level jobs is critical to the development of the skills needed by higher-level, more advanced workers. The only way people develop the expertise to do high-level jobs is to start in entry-level jobs. And the only way people become good general managers in companies that provide a variety of goods and services is by their working in lower-level jobs in different parts of the company. So my concern is that the rapid advance of AI will undermine the training that creates highly skilled people.
(This is also why I, as a former college teacher, worry about college students writing papers with AI. Someone who never learns to write a good five-page paper without assistance will never develop the skills to write a novel or a scientific treatise or a powerful work of philosophy, let alone a strategic plan for a business or political campaign.)
One might expect that individual businesses would be aware of this problem and would thus sustain and even subsidize entry-level positions with the goal of ensuring that they have a qualified high-level workforce in the future. But in an economy in which workers are so mobile—in which they change jobs and move around the country so quickly—a collective action problem may arise. Because individual businesses cannot assume that they will benefit from providing entry-level training to workers, they are likely to underinvest in it. The result is that government is likely to have to step in to require and subsidize these jobs—or the necessary training that replaces them—just as it does to subsidize education, including higher education, for similar reasons. And as is the case for education as a whole, government subsidies of entry-level training positions will be important not only for the economic health of the economy but for providing equality of opportunity to young people. Given how bad the United States has been at funding education and training at all levels, there is good reason to worry about this problem.
There is also reason to be concerned that if AI replaces more lower-level jobs than upper-level jobs, this shift will reduce the rewards to the former and increase the rewards to the latter, thus leading to greater generational inequality. This is one way AI could lead to the economy generating insufficient demand to ensure full employment and high levels of investment. It’s also possible the insufficient support for people transitioning from old to new jobs will also lead to insufficient demand. In both these ways AI could create the kind of secular stagnation that inequality always seems to generate in market economies. And if younger people are bearing the brunt of this inequality, this may reduce their ability and inclination to have children. Progressive taxation that pays for redistribution programs that focus on family supports will be necessary to overcome these difficulties.
Another form of redistribution may be necessary to ensure that everyone can reap the benefits of the individualization in such things as medicine, therapy, education, and other social services that, as I’ve pointed out a few times, can be one of the great benefits of AI. An economy in which those with low incomes receive generic education or medical treatments while those with higher incomes receive far more effective individualized services is deeply troubling. A serious commitment to social justice will require attention to these issues.
Conclusion
In this piece I’ve made the best case I can for the thesis that we should not be worried about AI eliminating all human jobs. I recognize that some of my claims may be questionable. To be honest, I worry that I may be underestimating how far AI can go in replacing human beings even in the service sectors like medicine, education, and therapy where the human ability to understand other people and creatively respond to them is so important. Yet, even if I am somewhat wrong there, the other arguments I put forward suggest that AI will not eliminate all human jobs quickly or at all.
But as the last section of this post shows, we will need to address some vital questions about public policies that surround the rollout of AI if the benefits of AI are widely shared. I should add that the last section of the post does not do much more than skim the surface of deep issues that we will need to address, and soon.
And, of course, I have not addressed at all the question of what kinds of taxes would be best to raise the revenues we will need to provide transition and retraining services, to ensure the survival of entry-level jobs, to guarantee that individualized care and treatment is available to all, and to ensure that AI does not create devastating intergenerational or other forms of inequality.
One question that we should address in the future is whether those revenues should be generated by general taxes on the income and wealth of our most well-off citizens. Or is there a strong case to institute taxes on AI itself so as to mitigate some of the potential problems I discuss here as well as other problems including the environmental