Will Robots Take Your Job?

All these jobs will be gone like tears in the rain

why does the robot look like roman in the dick pic scene from succession

A pointsman’s back straightened itself upright suddenly against a tramway standard by Mr. Bloom’s window. Couldn’t they invent something automatic so that the wheel itself much handier? Well but that fellow would lose his job then? Well but then another fellow would get a job making the new invention?

James Joyce, “Ulysses”

Large Language Models, AI, and other newfangled technologies have shown great promise: for art, programming, writing, etc. This has people marveling at the potential benefits, but also at the costs: what are we going to automate next?

Do android economists use electric supply and demand?

Concerns over automation may seem new, but they’re not. All the way back to the Luddite movement of the 19th century to today, history is plagued by anti-machine fears. For example, concerns were also really strong during the 1950s and 1960s, to the point where LBJ put together a blue ribbon panel on this exact issue. Humans becoming as employable as horses doesn’t sound pretty - but is it going to happen?

The most important thing to keep in mind is that jobs aren’t automated. Tasks are automated. When all of the tasks in a job are automated, that job disappears, and the people who hold that job eventually lose it to robots. Something relevant is that the counterpart of a task is a skill - the thing you need to complete the task. For instance: driver is the job, driving is a task, and driving is a skill - but there’s more tasks that require other skills, like fueling the tank or loading stuff into a car. Note that both of those are really big caveats. A lot of jobs can be partially automated, but not fully. When people say “X% of jobs are at risk of being automated”, what they mean is “X% of jobs contain tasks that could be automated”.

Does that mean all of these jobs will be automated? Let’s look at a real world example: ATMs. Before ATMs were invented, people went to the bank and did all of their operations with real people - withdrawals, deposits, plus all of the clerical stuff. After ATMs were invented, some new jobs were created (say, maintaining the machines) but many of the bank clerks lost their job. Except they didn’t actually - because going to the bank was less of a pain, people started going more often, which meant banks had to open more branches. And since a lot of bank work was still done by person, bank employment increased after a decade or two.

A tale of two cities

The ATM story points to something crucial: when technology automates tasks, the skills required by new tasks don’t stay the same. For instance, the new bank employees stopped needing skills at “counting money” and started needing skills more in line with customer service. This meant that many people who got laid off might not have gotten rehired by banks, because they didn’t have the new necessary set of skills.

Economists call this phenomenon skill-biased technical change, i.e. a shift from one group of skills (say, assembling things) to a different one (operating assembling machines) due to a new technology. Skill biased technical change (SKBTC) points to one simple fact: workers who can do jobs that benefit from automation of tasks gain, while those who lose out lose - and there’s no reason why the losses and the gains have to be spread out evenly. It also, however, doesn’t mean that the workers themselves have to lose out by losing their jobs. For example, automation allowed the share of workers in US agriculture to fall from 40% in 1900 to 2% just a hundred years later. This didn’t mean that, on net, jobs were lost - workers just moved to cities and got work in manufacturing, transportation, or (later) services.

In general, there’s three groups and three patterns: jobs that are either highly intensive in education and highly paid (stuff like professional labor, or management), jobs that highly intensive in social/physical labor and poorly paid (say, personal care, or food services), plus a “middle class” of jobs”, which are somewhere in the middle for both (sales, clerical work, etc). The first two have grown, whereas the last group shrank - was it because of changes in technology?

This phenomenon, known as labor polarization, is widely acknowledged: “low skill” manual jobs have declined in terms of relative earnings, while higher skill jobs have increased, while the middle was hollowed out. This is common to most advanced economies, and has been happening for a pretty long time. A not entirely unpredictable consequence of labor polarization is that inequality has risen in a variety of dimensions. This isn’t all bad, though: it has helped narrow the gender wage gap. The growing divergence between the educational attainment of men and women have resulted in blue-collar work becoming disproportionately male, while white collar work gets disproportionately female. However, this doesn’t explain the causes of polarization.

A first factor could simply be deunionization: labor unions helped protect low skill wages, but since union density declined in the US, low skill wages declined as well. Secondly, a reason for why skilled labor is in higher demand than unskilled labor (until at least the pandemic) was technology, obviously (otherwise why bring it up). “Middle skill” jobs were easier to automate, because they contained mostly repetitive tasks, such as doing math - aka being a “computer”. However, this is controversial: the skill biased technical change explanation doesn’t really account for a lot of the change in polarization, per many metrics. Also, the fact that wages are low at the low skilled level might discourage automation, rather than be a result of it. So it’s possible that the change in the skills composition of the workforce helps explain most of the shift on its own: the interactions between skills, tasks, and jobs are pretty complex.

So it seems that, so far, it’s not very likely that there emerges a class of tech workers making 700k while everyone else slums it out. Similarly, the fact that low skill wages have rapidly increased in recent years puts another big damper on the SBTC train.

It’s more fun to compute

The first half of this post, mainly, was written over two years ago, and then I stopped because I lost interest in the topic and, later, because of a simpler reason: ChatGPT. It was a very new technology when it first came out, meaning that I had to wait around to see what it could do before having a take. And I don’t use it, but I generally have a decent grasp on what it can do now.

The common take on ChatGPT and other LLM (“large language model”) technology is that it will cut into the demand for higher-end, more “cognitive” work - art, bookkeeping, consulting, design, etc. What, exactly, would that be like?

Paul Krugman, quite infamously, announced that “By 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s.” Besides the fact that, per the actual bounds of the debate in the late 1990s, he was right, or that the fax machine was in fact a pretty big deal, he also got to something important: the internet merely modified the economy, but didn’t fundamentally transform it like originally predicted¹.

The key idea to comprehend is that of the productivity S curve: the economic growth caused by any given technological invention, no matter how large, has a small plateau, a swift takeoff and then another plateau. This concept is key, for some, to understanding why the growth of all major economies plateau’d in the 2000s and 2010s: at the same time as demand sagged, population growth crashed, and, crucially, technological improvements simply provided less bang for their buck. A problem with this view, of course, is that a lot of the investment required to use certain types of large-scale technologies is very difficult to measure, putting this interpretation into question - perhaps there is simply no technological slowdown².

Well, we actually have a test case: computers. Looking at how computers affected the labor market, common sense would dictate that they would have replaced many educated and skilled workers. In fact, quite the opposite occurred: the demand for college graduates grew more rapidly than for other types of employees, leading to an increase in the amount of college graduates as well. This also happened within industries, which means that “jobs” were not actually destroyed, but rather, shifted around. The occupations that don’t use computers experienced job losses, in fact.

The main, most obvious reason for this is that jobs are made up of tasks, and that those tasks require certain skills to be performed. If those skills are complemented by new technologies, then the job holder is (seemingly paradoxically) benefitted by automation, not harmed. It’s also a consistent trend that experts, journalists, and commentators are remarkably bad at knowing how much of a job can be automated away by new tools, or whether the margin of adjustment will be to make certain jobs cheaper or to make fewer of them. Obviously, this has clear distributional impacts, and can be affected by market power and other such imperfections. Institutions in the labor market, training levels, and education policy can shape the workforce to be more or less complimentary with technology. And there are also other possible benefits, since technology can make skills and experience more legible to employers, or make connecting with new workplaces more straightforward.

So this means that there might not be short-term job loss, but what about long term? Wouldn’t people simply get too expensive, or too bad? Not really. Economic value is socially constructed, by which we mean the economy is made up of people doing things for other people. The most obvious such example is the service sector: a massive chunk of the economy is people doing in person labor for other people, in ways that are extraordinarily difficult to automate. Japan, for instance, attempted to automate the elder care sector, and it was a disaster: robots required constant maintenance and were not a good replacement in terms of actual performance. There’s also the fact that people are constantly coming up with new ways to spend their time: influencers, yoga teachers, youtubers, stand up comedians, etc.

But wouldn’t wages go down if everyone worked with old people and children? Not precisely. The key to answering this issue is a concept known as cost disease: the wages in all industries, not just in high-tech ones, is positively correlated to the most advanced sectors. Why? Imagine you’re a violinist. You can’t really improve your productivity at playing music. But you can, actually, choose to stop playing music and get a boring 9 to 5 filling spreadsheets if your wages, as a musician, don’t increase. Which means that even the least productive sectors see pay rises when productivity in other sectors increases, lest their employees leave and do something else.

The last concern, naturally, is short-term job losses: so the concept of “work” is safe in the long term, and the general gist of it ist that it’s sorted out in the medium term. But what about the adjustment? Wouldn’t mass unemployment be bad? Well, yes. But the thing is, mass unemployment is not a technological problem; it’s a macroeconomic problem - in fact, it’s the source of so many errors in economic reasoning. And while the labor market for coders or artists or Japanese nurses might not balance itself out, the Labor Market as a whole does have someone looking over it: the Central Bank. To qutoe Paul Krugman:

in reality the Federal Reserve Board actively manages interest rates, pushing them down when it thinks employment is too low and raising them when it thinks the economy is overheating. You may quarrel with the Fed chairman's judgment--you may think that he should keep the economy on a looser rein--but you can hardly dispute his power. Indeed, if you want a simple model for predicting the unemployment rate in the United States over the next few years, here it is: It will be what [Powell]wants it to be, plus or minus a random error reflecting the fact that he is not quite God.

Conclusion

In quite a few years (…) we may be able to perform all the operations of agriculture, mining, and manufacture with a quarter of the human effort to which we have been accustomed. (…) We are being afflicted with a new disease of which some readers may not yet have heard the name, but of which they will hear a great deal in the years to come--namely, technological unemployment. (…) But this is only a temporary phase of maladjustment. All this means in the long run that mankind is solving its economic problem.

John Maynard Keynes, (1930) “Economic Possibilities for Our Grandchildren”

I started writing this blog post two years ago (see: the DEVO reference) and the ideas of interest were very very different. But in general, I came to the same conclusions: automation might shake up some professions, perhaps profoundly, but ultimately they won’t abolish work OR make society some sort of dystopian hellscape.

The main issue here isn’t really that any specific technologies would drastically alter the entire economy, because they simply won’t. People will still have to work doing something. The issue is more “what would that something be” and “how would that something be distributed” - the problems with automation don’t come with the technologies themselves, but rather, with the politics and macroeconomics of the adjustments to be made. It’s always been like that!

1. The benefits of the internet, as evidenced by you being able to read my blog, are actually probably quite large.

2. There’s also the view that the economic slowdown isn’t real or that it is in fact good.