In the Long Run

My problem with Robert D. Atkinson’s op-ed in the Wall Street Journal on the likely impact of artificial intelligence on jobs and our daily life:

Anthropic CEO Dario Amodei said last week that artificial intelligence could eliminate half of all entry-level white-collar jobs within five years and cause unemployment to skyrocket to as high as 20%.

He should know better—as should many other serious academics, who have been warning for years that AI will mean the end of employment as we know it. In 2013 Carl Benedikt Frey and Michael A. Osborne of Oxford University produced a research paper estimating that 47% of U.S. employment was at risk of being eliminated by new technologies.

Before we resign ourselves to obsolescence at the hands of our new robot overlords, we’d do well to recognize that humans have experienced technological disruptions before, and we adapted to meet them. AI won’t be any different.

In the first half of the 20th century, tens of thousands of men and boys across America worked as pinsetters in bowling alleys. In 1946 AMF introduced an automatic pin-setting machine, and by the mid-1950s those jobs were mostly gone. Similarly, there was a time when the elevators in hotels and office buildings across the country were staffed by human operators. In the 1920s and ’30s, elevator companies began installing “robot elevators” with automatic controls, and eventually elevator operators all but disappeared.

Data from the Census Bureau’s American Community Survey tell countless other similar stories—from the decline of agricultural field workers due to motorized tractors to the rise and fall of “motion picture projectionists,” who operated projectors in movie theaters. Entire categories of jobs were wiped out, yet automation has never created a mass lumpenproletariat.

AI doomsayers frequently succumb to what economists call the “lump of labor” fallacy: the idea that there is a limited amount of work to be done, and if a job is eliminated, it’s gone for good. This fails to account for second-order effects, whereby the saving from increased productivity is recycled back into the economy in the form of higher wages, higher profits and reduced prices. This creates new demand that in turn creates new jobs. Some of these are entirely new occupations, such as “content creator assistant,” but others are existing jobs that are in higher demand now that people have more money to spend—for example, personal trainers.

is that I agree with him. I think that in the long run artificial intelligence will increase productivity and create new jobs that we can’t even imagine today. There are, however, some problems with the rosy scenario that he paints. Chief among those are timeframe and that the benefits of that automation will not be distributed evenly through the population.

We shouldn’t forget John Maynard Keynes’s wry warning: in the long run we’re all dead. Although in the long run I deeply agree that this new automation will increase productivity and economic welfare, in the short term it is likely to cause enormous dislocation.

The first sector in which that is likely to happen is information technology. Indeed, it is happening right now. Not only are their large numbers of layoffs but there is tremendous uncertainty. Neither workers nor even their managers have any real idea of what will happen. I suspect that when the dust has settled it will be much like previous such developments—the effect of AI will be to make developers more productive rather than to eliminate developers.

The examples Mr. Atkinson provides are terrible: agricultural stoop laborers, elevator operators, insurance customer service. The first two are unskilled work and I would be willing to bet a shiny new dime (how long until dimes are abolished, too?) that AI is not deployed to replace insurance customer service workers for the foreseeable future—they’ve already largely been replaced by offshore call center workers and autoresponse systems.

The real potential of generative AI is in making highly skilled workers much more productive and effective—physicians, lawyers, accountants, and so on. Whether GAI will be allowed to do that is another question. I suspect that physicians, lawyers, etc. will resist such efforts strongly and, since they control what is allowed to be done, it’s an open question.

That brings me to another issue: who will capture the economic surplus provided by more efficient artificial intelligence? Is there a single investor anywhere who wants to invest in AI to increase workers’ wages? Or is the motivation to increase the compensation of those at the top of the economic pyramid? I think it’s obvious.

That brings me to my final observation. Mr. Atkinson observes:

Further, there is a great deal of work that only humans can do. Self-driving school buses will still need an adult to watch the kids. As for police, AI robots won’t be arresting criminals anytime soon. It’s a similar story for fish and game wardens, fashion models, priests, stonemasons, plumbers and flight attendants. Most occupations involve working with other people, with things or with ideas that are too complex for AI to handle alone. People in the last category include legislators, CEOs, antitrust attorneys and so on.

I think that’s almost the diametric opposite of reality. There are lots of jobs that are too complex for people to handle without AI. I would also mention that none of the examples he mentions (fish and game wardens, fashion models, etc.) are growth occupations. There are fewer of all of those than there were 50 years ago.