The Real Economic Impact of Artificial Intelligence (AI): Hint It Isn’t Massive Unemployment
In progressive and tech circles, there is a lot of discussion about the impending doom of AI (artificial intelligence) on the economy. They expect massive job losses as tasks done by people today are automated away by machines. Indeed, this conversation has become the central issue behind some 2020 candidates like Andrew Yang and a unifying force among libertarians and progressives for a universal basic income.
A lot of the fear in this discussion focuses on technology that does not actually exist today, looking at general intelligence robots that can do everything a human can. That is far away from where AI technology is today, which is mostly driven by advances in machine learning, and where it can reasonably be expected to go in the near future.
In machine learning, there are roughly two types of problems that computers try to solve: unsupervised learning and supervised learning. Most valuable applications today are from supervised learning, which answers questions like how to predict some variable (Y) using a bunch of input data features (X).
Supervised learning (of which the hottest buzz is in deep learning, a specific type of supervised model) can be used to answer questions like “which of my customers are most likely to churn?” or “what product would this customer be most likely to purchase?”. These models are only really as good as the data put into them, but they are able to improve over time as more data are collected, and the model sees more examples.
When we look to forecast the impact that AI (and therefore mainly applications of supervised learning) can have on the economy, it is therefore limited in scope to problems of the form “predict Y from X”, baring unexpected breakthroughs. The space of tasks this can effect and automate today are things that it takes humans around a second or less to do, and that time window will increase as advances in AI/machine learning are made. That limited scope of tasks is not the dooms day scenario of taking away all jobs that many expect.
This is not to say that AI will not fundamentally restructure the economy, but the effect won’t be massive unemployment. The real impact, like previous examples of skill biased technological change, will be to select for certain skills over others as the technology creates new jobs and new opportunities, while automating other tasks.
To see this we can look at past instances of massive technological change as case studies of what to expect from AI.
Across decades, from the printing press to calculators to robotic arms, people were continuously worried that technology would take away their jobs. What happened instead was that new technology diminished the need for certain skills/occupations while increasing the need for others that could either (1) work with the new technology, (2) supply the new technology with the necessary inputs, (3) use the products from new technology in other downstream applications, or (4) satisfy demand of people who made money with the new technology.
A great example is the invention of the spreadsheet and software like Microsoft Excel that automated a lot of record keeping in the late 1970s/early 1980s. As the Wall Street Journal notes, bookkeeping jobs plummeted as the software replaced the manual tasks people were previously required to do. Meanwhile demand for analysts and accountants soared as companies were now able to keep track of more data and needed help analyzing and interpreting it. Right there we see how new technology differentially selected for certain skills, increasing their demand over others, while the skills replaced saw their value in the labor market decline.
We can see the exact mechanisms behind this through a more systematic analysis from a paper by Autor and Salomons, two economists. They measure the relationship between various metrics and the growth in something called TFP (total factor productivity) over time. TFP is a common measure of technological change in the economics literature that strongly correlates with things like the number of patents (it is essentially the parts of economic growth that cannot be explained by labor and capital/machines).
A new technology is deemed labor displacing if it reduces the share of income that goes to workers and instead increases the share that goes to capital/machines. They also look at the relationship between TFP growth and employment and wages.
We should note that an increase in labor’s share of income technically means that the increase in the economy from new technology is disproportionately going to capital over labor, but it doesn’t mean that labor itself does not benefit; rather, it means that labor is benefitting less from new technology than capital, driving its overall share lower.
They break down the impact of new technology on these outcomes into four parts: (1) direct effects, (2) final demand effects, (3) upstream effects, and (4) downstream effects.
Obviously, within an industry, jobs focused on doing tasks that machines automate go away, but as the industry overall becomes more productive due to the new technology, labor demand in non-automated tasks could actually increase. This is the (1) direct effect that Autor and Solomans term the “Uber” effect. Technological improvement in an industry can raise both labor productivity and employment within the impacted industry.
The ride app sector is an example of this because while Uber’s platform automated the matching of riders to drivers, removing that task from one humans would do, it raised both productivity and demand for drivers, increasing demand for those jobs.
The (2) final demand effect is the increase in the economy wide demand from new technology. Owners of capital for new technology benefit from renting their equipment to companies, and that rise in income raises their demand for purchasing goods and services, which in turn drives demand for jobs involved in those areas, creating new opportunities for labor.
As a basic example, the internet created a whole host of millionaires and billionaires through founding and/or working for tech companies, leading to surges in demand for goods and services those people liked even though those companies created products that automated away lots of existing job tasks (like Microsoft with Excel).
The (3) Upstream and (4) downstream effects are how TFP growth impacts supplier and customer industries.
New technology can reduce the cost of production in one input, leading to a rise in demand for other inputs. As an example improvements in lithium ion batteries created opportunities to leverage them in smaller, mobile devices, allowing the creation of smartphones. This led to an increase in demand for microchips, an upstream supplier industry, as they are used in phone production, spawning new opportunities for jobs in that area. Similarly, the creation of mobile phones spawned a host of companies developing apps for them, driving demand in customer industries as well.
The net balance then of technology on labor involves analyzing these four different impact channels: (1) direct effects, (2) final demand effects, (3) upstream effects, and (4) downstream effects.
When Autor and Solomons crunch the numbers, they find that technology is typically net negative in (1) direct effects as the new technology displaces labor in industries where it is applied, but the other channels tend to be overall positive on labor markets and outweigh the direct negative impact on the industry itself.
This means that generally technology is actually labor augmenting, increasing employment, but it creates opportunities in different places. Because it creates opportunities in different industries, it selects for different skills than those who lose their jobs because of technology have.
In this way the introduction of new technology is quite similar to the impact of globalization: it has concentrated harm on a few individuals who will lose the jobs they have today, but the overall impact on the economy is larger as it creates new economic opportunities for people and results in generally higher quality products and/or lower prices.
Applying these lessons from past technology shocks to a popular topic today, we can see the real expected impact of AI on truck drivers. While even the driving of trucks might become automated, there will still be a role for humans to oversee the transport of cargo, unload it, refuel the truck, and many other potential tasks akin to how pilots are still necessary in plans even with autopilot.
The central policy question then is not whether we can support massive amounts of unemployment as all jobs disappear due to improvements in AI, but rather, can we do a better job of helping people adjust to the inevitable disruption to labor markets from technology, learning lessons from the first days of globalization and first wave of automation that this adjustment cost is not frictionless and can have a huge negative impact on the people who are directly affected.
We should push for policies that can aid this adjustment like wage insurance, higher minimum wages, reductions in occupational licensing, and a host of other improvements to make our labor markets more dynamic and responsive to structural changes.