Why predictions of an imminent economic revolution are overstated
Excerpt – [..] in the 1960s Robert Fogel published work about America’s railways that would later win him a Nobel Prize in economics. Many thought that rail transformed America’s prospects, turning an agricultural society into an industrial powerhouse. In fact, it had a very modest impact, Fogel found, because it replaced technology—such as canals—that would have done just about as good a job. The level of per-person income that America achieved by January 1st 1890 would have been reached by March 31st 1890 if railways had never been invented.
Of course, no one can predict with any certainty where a technology as fundamentally unpredictable as AI will take humans. Runaway growth is not impossible; nor is technological stagnation. But you can still think through the possibilities. And, so far at least, it seems as though Fogel’s railways are likely to be a useful blueprint. Consider three broad areas: monopolies, labour markets and productivity. [..]
A new technology sometimes creates a small group of people with vast economic power. John D. Rockefeller won out with oil refining and Henry Ford with cars. Today Jeff Bezos and Mark Zuckerberg are pretty dominant thanks to tech.
[..] Monopolies often arise when an industry has high fixed costs or when it is hard to switch to competitors. Customers had no alternative to Rockefeller’s oil, for instance, and could not produce their own. Generative AI has some monopolistic characteristics. GPT-4, one of OpenAI’s chatbots, reportedly cost more than $100m to train, a sum few firms have lying around. There is also a lot of proprietary knowledge about data for training the models, not to mention user feedback.
There is, however, little chance of a single company bestriding the entire industry. More likely is that a modest number of big firms compete with one another, as happens in aviation, groceries and search engines. No AI product is truly unique since all use similar models. This makes it easier for a customer to switch from one to another. The computing power behind the models is also fairly generic. Much of the code, as well as tips and tricks, is freely available online, meaning that amateurs can produce their own models—often with strikingly good results.
[..] In the past decade the average rich-world unemployment rate has roughly halved. The share of working-age people in employment is at an all-time high. Countries with the highest rates of automation and robotics, such as Japan, Singapore and South Korea, have the least unemployment. A recent study by America’s Bureau of Labour Statistics found that in recent years jobs classified as “at risk” from new technologies “did not exhibit any general tendency toward notably rapid job loss”. Evidence for “hollowing out” is mixed. Measures of job satisfaction rose during the 2010s. For most of the past decade the poorest Americans have seen faster wage growth than the richest ones.
[..] history suggests job destruction happens far more slowly. The automated telephone switching system—a replacement for human operators—was invented in 1892. It took until 1921 for the Bell System to install their first fully automated office. Even after this milestone, the number of American telephone operators continued to grow, peaking in the mid-20th century at around 350,000. The occupation did not (mostly) disappear until the 1980s, nine decades after automation was invented. AI will take less than 90 years to sweep the labour market: LLMs are easy to use, and many experts are astonished by the speed at which the general public has incorporated ChatGPT into their lives. But reasons for the slow adoption of technology in workplaces will also apply this time around.
In a recent essay Mark Andreessen of Andreessen Horowitz outlined some of them. His argument focuses on regulation. In bits of the economy with heavy state involvement, such as education and health care, technological change tends to be pitifully slow. The absence of competitive pressure blunts incentives to improve. Governments may also have public-policy goals, such as maximising employment levels, which are inconsistent with improved efficiency. These industries are also more likely to be unionised—and unions are good at preventing job losses.
Modest labour-market effects are likely to translate into a modest impact on productivity—the third factor. Adoption of electricity in factories and households began in America towards the end of the 19th century. Yet there was no productivity boom until the end of the first world war. The personal computer was invented in the 1970s. This time the productivity boom followed more quickly—but it still felt slow at the time. In 1987 Robert Solow, an economist, famously declared that the computer age was “everywhere except for the productivity statistics”.
The world is still waiting for a productivity surge linked to recent innovations. Smartphones have been in widespread use for a decade, billions of people have access to superfast internet and many workers now shift between the office and home as it suits them. Official surveys show that well over a tenth of American employees already work at firms using AI of some kind, while unofficial surveys point to even higher numbers. Still, though, global productivity growth remains weak.
AI could eventually make some industries vastly more productive. A paper by Erik Brynjolfsson of Stanford University and colleagues examines customer-support agents. Access to an AI tool raises the number of issues resolved each hour by 14% on average. Researchers themselves could also become more efficient: GPT-X may give them an unlimited number of almost-free research assistants. Others hope AI will eliminate administrative inefficiencies in health care, reducing costs.
But there are many things beyond the reach of AI. Blue-collar work, such as construction and farming, which accounts for about 20% of rich-world GDP, is one example. An LLM is of little use to someone picking asparagus. It could be of some use to a plumber fixing a leaky tap: a widget could recognise the tap, diagnose the fault and advise on fixes. Ultimately, though, the plumber still has to do the physical work. So it is hard to imagine that, in a few years’ time, blue-collar work is going to be much more productive than it is now. The same goes for industries where human-to-human contact is an inherent part of the service, such as hospitality and medical care.
AI also cannot do anything about the biggest thing holding back rich-world productivity growth: misfiring planning systems. When the size of cities is constrained and housing costs are high, people cannot live and work where they are most efficient. No matter how many brilliant new ideas your society may have, they are functionally useless if you cannot build them in a timely manner. It is up to governments to defang NIMBYs. Technology is neither here nor there. The same goes for energy, where permitting and infrastructure are what keep costs uncomfortably high. [..]
AI may change the world in ways that today are impossible to imagine. But this is not quite the same thing as turning the economy upside down. Fogel wrote that his argument was “aimed not at refuting the view that the railroad played a decisive role in American development during the 19th century, but rather at demonstrating that the empirical base on which this view rests is not nearly so substantial as is usually presumed”.
Full article, The Economist, 2023.5.7