My favorite analogy for understanding technological unemployment is “peak horse.” Below is the version told in Greg Clark’s book A Farewell to Alms:
there was a type of employee at the beginning of the Industrial Revolution whose job and livelihood largely vanished in the early twentieth century. This was the horse. The population of working horses actually peaked in England long after the Industrial Revolution, in 1901, when 3.25 million were at work. Though they had been replaced by rail for long-distance haulage and by steam engines for driving machinery, they still plowed fields, hauled wagons and carriages short distances, pulled boats on the canals, toiled in the pits, and carried armies into battle. But the arrival of the internal combustion engine in the late nineteenth century rapidly displaced these workers, so that by 1924 there were fewer than two million. There was always a wage at which all these horses could have remained employed. But that wage was so low that it did not pay for their feed.
Yes. Horses are not people. But if you swap the word “horse” for “people”, many apparently compelling arguments for or against technological unemployment reveal their flaws. For example see Brad Delong taking Joel Mokyr somewhat to task for not taking on peak horse. Peak horse clarifies how technological unemployment works. In the medium term I think these fears are overblown, but ultimately there’s a certain logic to it, as we’ll see.
One old (discredited) argument for technological unemployment, the lump of labor fallacy, presumes there’s only so much work to do. This makes sense, but only if human desires are finite. Good luck with that. A new and better version acknowledges past concerns were wrong, but says computer technology is different. Computers and robots don’t just replace physical labor, they replace cognitively demanding labor. This second version is far more difficult to argue against. To be clear, a third version argues technological change drives short term unemployment. You learned one job, then new technology makes your job obsolete. So you have to learn something new or become unemployed. As Erik Brynjolfsson and Andrew McAfee point out in their book The Second Machine Age, the exponential progress of Moore’s Law has made this technology skill set mismatch problem much worse. That’s bad. And there’s a strong consensus this happens. But this effect doesn’t drive permanent technological unemployment, which is what people are now starting to fear.
Below is our horse analogy in stylized form.
Technology improves over time, moving left to right. “High skill” horses at top in green are mostly employed. “Low skill” horses in red are mostly unemployed. When cars are invented, the skill threshold required for horse employment rises. Cars substitute for “low skill” horses, leaving them unemployed.
If we generalize this to people, we expect the lowest skilled workers will find it harder and harder to get jobs over time. Let’s call this Model 1: tech and robots substitute for people.
In contrast, maybe what separates people from horses is they participate in a market economy. So there is a threshold at play, but one completely different from above. This threshold is fixed, having to do with the minimal set of social and cognitive skills needed to participate in labor markets. As long as someone is above that fixed lower bound (not handicapped or disabled), improving technology doesn’t matter for employability either way. And because of comparative advantage, even if robots become cognitively sophisticated, people can still trade with them by shifting their effort to work on things they are relatively better at. So in this model we expect the low skill unemployment rate to be independent of improvements in technology over time. Let’s call this Model 2: tech and robots complement people.
In economic terms of course these two models contrast complementary versus substitute goods. Schematically as below:
The biggest distinction between these two stylized models is in what they predict happens to low skill workers over time. If the minimal skill level needed for market participation continually rises as technology improves, then substitute model 1 applies. If the minimal skill level appears flat, largely independent of technology improvements, then complement model 2 applies.The fate of the lowest skill marginal workers tells the tale. Also note there’s nothing in these models logically tied to the computer age. So if we want we can select a larger historical time period (say a few hundred years) in judging their applicability.
Of course, as always in social science, there’s confounding factors. As technology improves, wealth grows and society can afford more charity. This puts a rising floor on the minimally acceptable wage. That’s great! But a rising floor on minimal wages may make the data appear to fit the substitute model, even if the underlying economics are driven by the complement model. A related point made by Tyler Cowen in Average is Over is computers greatly increase income stratification. I think that’s true. Though the stronger argument for technological unemployment requires computers to be able to do all conceivable types of work low skill humans can do, but more cheaply. That is, we’re back to substitution. For more thoughts on Cowen’s book see my post.
When I searched the New York Times on computers and the disabled, I found this 2006 story Computer Technology Opens a World of Work to Disabled People. Quote:
Such arrangements are bringing jobs to thousands of people with disabilities, including those with spinal cord injuries and vision loss. Fast computers and broadband connections have become so inexpensive and reliable that location is now not an issue for certain jobs, like customer service.
At the same time, an abundance of technology is available to help disabled people operate computers, like software that lets a blind person use a keyboard instead of a mouse to navigate a program, and voice synthesizers that turn text into speech. There are also alternatives to the mouse for people with limited use of their arms.
When I took a quick look at Social Security data on employment levels for people with disabilities, no surprise the data didn’t seem definitive either way. For what it’s worth, while both models apply to some extent, I find the complement model far more compelling. Certainly from a historical point of view, the minimal level of raw cognitive competence needed to participate in the market has not appeared to have changed much in two hundreds years. Yes, the kind of training and schooling needed now differs radically from the past. But one could just as easily argue (as the New York Times story does above) that improving technology has lowered the bar for market entry for people with disabilities.
One difficulty with the substitute model is it implies people are very inflexible (like horses), being unable to shift into new work which is complementary to what computers are good at. At an exaggerated level, the substitute model means Ricardo was wrong, as high skill and low skill won’t even bother to trade since low skill will sit around being technologically unemployed. Bottom line: I don’t find the new technological unemployment argument about computers any more compelling than the older ones.
But that’s only in the medium term. A time when artificial intelligence is nascent, or at least not ubiquitous. What happens in a few hundred or even a thousand years? That’s where the picture is less clear, and peak horse may turn out to be right after all. For this kind of future forecast, let’s turn to Robin Hanson‘s 1998 paper Economic Growth Given Machine Intelligence. Quote: “Let us now take a standard neo-classical (Solow-Swan) growth model with diminishing returns, and modify it minimally to include ordinary computers and machine intelligence.” And so:
Machines complement human labor when they become more productive at the jobs they perform, but machines also substitute for human labor by taking over human jobs. At first, expensive hardware and software does only the few jobs where computers have the strongest advantage over humans. Eventually, computers do most jobs. At first, complementary effects dominate, and human wages rise with computer productivity. But eventually substitution can dominate, making wages fall as fast as computer prices now do.
That is to say, at first computers and robots make us richer through complement effects. Likely much much richer. Something I would argue is happening today. Eventually robots become extremely good, drop to nearly zero cost, and are endlessly duplicated. So they work for next to nothing, taking away all human jobs. Substitute effects dominate. Peak human. The power of this argument lies in how it makes very few assumptions. It just projects standard economics forward, coupling it to a continuing projected decrease in the cost of computer hardware and software. Plus of course having eventual success in creating machine intelligence. How close we are is a matter of conjecture. But I’m certain this scenario of ubiquitous, nearly costless robots is not what most people have in mind when they worry about technological unemployment. Even though it seems difficult to make a strong case for robot substitution without it.
When this eventually does start to happen, I hope someone has read their era’s version of Thomas Piketty. Call it Capital in the Thirty-First Century. That way humanity will own enough capital in the explosively growing robot empire to allow our great-grandchildren to retire in style.