The era of “generative†artificial intelligence has truly arrived. OpenAI’s chatbots, which use the technology of large language models, gave the go-ahead in November. Now hardly a day goes by without some spectacular breakthrough. The music industry was recently rocked by an AI-created song featuring a fake Drake and The Weeknd. Programs that convert text to video are creating some pretty compelling content. Before long, consumer products like Expedia, Instacart, and OpenTable will connect to OpenAI bots, allowing you to order food or book a vacation by typing text in a box. According to a recently leaked presentation by a Google engineer, the tech giant appears to be concerned about the ease with which its rivals are making progress. And we are going to see more things; probably many more.
The development of artificial intelligence raises profound questions. Although perhaps the one that stands out the most is simple. What does it mean for the economy? Many have high expectations. New research from Goldman Sachs bank indicates that “the widespread adoption of artificial intelligence could lead to a 7% increase, or almost $7 trillion, in annual global GDP over a ten-year period.” Academic studies point to a three percentage point increase in annual labor productivity growth for companies that adopt the technology, which would represent a huge increase in revenue added up over many years. An analysis published in 2021 by Tom Davidson of Open Philanthropy, a grantmaking organization, puts the odds of “explosive growth†(defined as an increase in global production of more than 30% per year) at some point in this century. Some economists imagine, only half jokingly, the possibility of world incomes becoming infinite.
However, financial markets point to much more modest results. Last year, the share prices of companies involved in artificial intelligence performed worse than the world average, although they have risen in recent months (see chart). Interest rates are another clue. If you believed that technology was going to make everyone rich tomorrow, rates would rise because there would be less need to save. Inflation-adjusted rates and consequent GDP growth are closely correlated, notes research by Basil Halperin of the Massachusetts Institute of Technology (MIT) and colleagues. However, since the AI ​​buzz began in November, long-term rates have fallen; and they remain very low in historical terms. Financial markets, the researchers conclude, “do not expect a high probability of…an AI-induced growth acceleration…at least over a 30-50 year time horizon.â€
To judge which group is right, it helps to look at the history of previous technological advances. It is something that helps investors. Because it’s hard to say that a single new technology has single-handedly changed the economy, for better or worse. Even the industrial revolution of the late eighteenth century, which many consider to be the result of the invention of the Jenny (multi-spindle) spinning machine, was actually due to a combination of all sorts of factors: the increased use of coal, the consolidation of property rights, the emergence of a scientific spirit, and many other things.
In the 1960s, Robert Fogel published a paper on American railroads that would later earn him the Nobel Prize in Economics. Many believed that the railroad had transformed America’s prospects, turning an agricultural society into an industrial powerhouse. In reality, it had a very modest influence because, as Fogel discovered, it took the place of technologies (such as channels) that would have done the same job well. The level of per capita income reached by the United States on January 1, 1890 would have been achieved on March 31 of that same had the railroad not been invented.
Of course, no one can predict for sure where a technology as inherently unpredictable as artificial intelligence will take humans. Runaway growth is not impossible; nor is technological stagnation. In any case, we can think of possibilities. And, for now at least, Fogel’s railways seem to make a useful model. Let’s consider three big areas: monopolies, labor markets, and productivity.
A new technology sometimes creates a small group of people with enormous economic power. John D. Rockefeller succeeded with oil refining and Henry Ford with automobiles. Today, Jeff Bezos and Mark Zuckerberg have achieved great dominance thanks to technology.
Many experts anticipate that the artificial intelligence industry will soon generate tremendous profits. Goldman analysts calculate in a recent paper that, under the most optimistic assumption, generative artificial intelligence could add about $430 billion to global annual revenue from enterprise software. The estimate is based on the purchase by the world’s 1.1 billion office workers of a few artificial intelligence programs at a total individual cost of about $400.
Any company would be delighted to capture a part of that money. Now, in macroeconomic terms, $430 billion is not a significant amount. Let’s assume that all revenue is converted to profit, which is unrealistic, and that all of those profits are made in the United States, which is a little more realistic. Even under these conditions, the relationship between corporate profits before taxes and GDP would go from the current 12% to 14%. Well above the long-term average, but not more than in the second quarter of 2021.
Those benefits could go to a single organization, perhaps OpenAI. Monopolies often arise when an industry has high fixed costs or when it is difficult to switch to the competition. Clients had no alternative to Rockefeller oil, for example; and neither could they produce it. Generative artificial intelligence has some monopolistic characteristics. GPT-4, one of OpenAI’s chatbots, is said to have cost more than $100 million to train, a sum available to few companies. There is also a lot of proprietary knowledge surrounding the data that models are trained on, not to mention user feedback.
However, there is little chance of a single company dominating the entire sector. Most likely, a small number of large companies will compete with each other, such as air transportation, supermarkets, and search engines. No AI product is truly unique, as they all use similar models. That makes it easier for a customer to switch from one to the other. The computing power behind the models is also quite generic. Much of the code, as well as tips and tricks, is freely available on the internet, which means that hobbyists can produce their own models, often with surprisingly good results.
“Currently, there don’t seem to be any systemic moats in generative artificial intelligence,†says a team at Andreessen Horowitz, a venture capital firm. The recent leak apparently coming from Google comes to a similar conclusion: “The barrier to entry for training and experimentation has been lowered from the full output of a large research organization to one person, one afternoon, and one powerful laptop.†There are already a few generative AI companies valued at more than $1 billion. The top corporate victor in the new age of AI isn’t even an AI company. At Nvidia, a computer company that makes hardware and software for artificial intelligence models, revenue from data centers is skyrocketing.
While generative artificial intelligence won’t create a new breed of rogue barons, that won’t be much comfort to many people. Many are concerned about their own financial prospects and, in particular, whether their job will disappear. There is no shortage of terrifying predictions. Tyna Eloundou of OpenAI and other colleagues have estimated that “about 80% of the US workforce could see at least 10% of their work tasks affected by the introduction of large language models.” Edward Felten of Princeton University and other colleagues perform a similar exercise. Legal services, accounting and travel agencies are among the professions most likely to lose out.
Economists have already made pessimistic forecasts in the past. In the 2000s, many feared the consequences of outsourcing on workers in the rich world. In 2013, two professors at the University of Oxford published a highly cited paper noting that automation could wipe out 47% of American jobs over the next decade or so. Others argued that, even without widespread unemployment, a “hollowout” would occur, with rewarding and well-paying jobs disappearing and their place taken by low-paying, mechanical functions.
What actually happened has caught everyone by surprise. In the last decade, the average unemployment rate in the rich world has roughly halved (see chart). The proportion of people of working age and employed is the highest in history. Countries with the highest rates of automation and robotics, such as Japan, Singapore and South Korea, have the least unemployment. According to a recent study by the US Bureau of Labor Statistics, in recent years jobs considered “at risk” from new technologies “did not show any overall trend toward remarkably rapid job losses.” The evidence for “emptying” is unclear. Indicators of job satisfaction rose in the 2010s. For most of the last decade, the poorest Americans have experienced faster wage growth than the richest.
This time, it could be different. The share price of Chegg, a company that offers home help, recently fell by half after admitting that ChatGPT was “having an impact on our growth rate of new customers.” IBM’s CEO has stated that the company plans to pause hiring jobs that could be replaced by artificial intelligence in the coming years. Now, are we facing the first signs of an imminent tsunami? Maybe not.
Suppose a job disappears when artificial intelligence automates more than 50% of its tasks. Or when workers are eliminated in the same proportion as the total number of automated tasks in the economy as a whole. In either case, according to Eloundou’s estimates, there would be a net loss of around 15% of US jobs. Some of them might move to sectors with a shortage of workers, such as the hospitality industry. In any case, there would undoubtedly be a large increase in the unemployment rate; in line, perhaps, with 15% briefly reached in the United States in 2020, during the worst of the covid-19 pandemic.
The problem with that scenario is that history teaches that job destruction occurs much more slowly. The automated telephone switching system (which replaced human operators) was invented in 1892. It took until 1921 for the Bell System to install its first fully automated office. Even after that milestone, the number of American manual trades continued to grow, peaking at about 350,000 workers in the mid-20th century. Employment didn’t disappear (mostly) until the 1980s, nine decades after automation was invented. It will take less than 90 years for artificial intelligence to spread through the job market: the great language models are easy to use, and many experts are amazed at the speed with which the general public has incorporated ChatGPT into their everyday lives. However, this time too, factors will work in favor of a slow adoption of technology in the workplace.
In a recent essay, Mark Andreessen of Andreessen Horowitz outlined some of them. Andreessen’s reasoning focuses on regulation. In sectors of the economy with large public intervention, such as education and health, technological change tends to be very slow. The absence of competitive pressure weakens the incentives for improvement. It is also possible that governments have public policy objectives (such as achieving maximum levels of employment) that are incompatible with improving efficiency. And it is very likely that those sectors are also unionized, and unions are good at preventing job losses.
Examples abound. Drivers on the public London Underground network are paid almost double the national average despite the fact that the technology to partially or fully replace them has existed for decades. Public bodies do not stop demanding over and over again that we submit forms with personal data on paper. In San Francisco, the world center of the rise of artificial intelligence, flesh-and-blood police officers are still employed to direct traffic at rush hour.
Many of the jobs threatened by artificial intelligence are in highly regulated industries. Let’s go back to Felten’s article from Princeton University. Fourteen of the 20 professions most exposed to artificial intelligence are teachers (foreign languages ​​rank near the top; geographers rank slightly stronger). However, only the most daring governments would replace them with artificial intelligence. Imagine the headlines. The same is true of police officers and artificial intelligence dedicated to fighting crime. The fact that Italy has temporarily blocked ChatGPT out of privacy concerns, and that France, Germany and Ireland are considering following suit, shows how concerned governments already are about the potentially job-destroying effects of artificial intelligence.
Perhaps governments will eventually allow the replacement of some jobs. However, the delay will allow room for the economy to do what it always does: create new kinds of jobs as others are eliminated. By lowering production costs, new technology can create more demand for goods and services, and boost jobs that are hard to automate. A paper published in 2020 by David Autor of MIT and other colleagues offered a surprising conclusion. About 60% of jobs in the United States did not exist in 1940. The job “manicurist†was added to the census in 2000. “Solar Photovoltaic Electrician†was added just five years ago. The AI ​​economy is likely to create new occupations that we can’t even imagine today.
The modest effects on the labor market are likely to translate into a modest impact on productivity, the third area. The adoption of electricity in factories and homes began in the United States in the late 19th century. However, there was no increase in productivity until the end of the First World War. The personal computer was invented in the 1970s. This time the increase in productivity was faster; although at the time it continued to seem slow. In 1987, economist Robert Solow declared that the computer age was “everywhere except productivity statistics.”
The world continues to expect a productivity boom linked to the latest innovations. Smartphones have been in widespread use for a decade, billions of people have access to high-speed Internet, and many workers now alternate between the office and home at their convenience. Official surveys indicate that more than a tenth of American employees already work at companies that use some form of artificial intelligence, and unofficial surveys point to even higher percentages. However, global productivity growth remains weak.
Over time, artificial intelligence could greatly increase the productivity of some industries. An article by Erik Brynjolfsson of Stanford University and colleagues discusses customer service agents. Access to an artificial intelligence tool increases the number of incidents resolved per hour by an average of 14%. The researchers themselves could also be more efficient: GPT-X could provide them with an unlimited number of research assistants almost free of charge. Others hope that artificial intelligence will eliminate administrative inefficiencies in healthcare, thereby reducing costs.
However, there are many things that fall outside the scope of artificial intelligence. Manual labor, such as construction and agriculture, which account for about 20% of the rich world’s GDP, is one example. A great language model is of little use to someone who collects asparagus. It could be a plumber fixing a leaky faucet: a widget would recognize the faucet, diagnose the fault, and advise solutions. However, the plumber will ultimately still have to do the physical work. So it’s hard to imagine that, in a few years, manual labor will be much more productive than it is now. The same is true of industries where human contact is an inherent part of the service, such as hospitality and healthcare.
Artificial intelligence can’t do anything either against the biggest brake on productivity growth in the rich world: failed planning systems. When cities are limited in size and housing costs are high, people cannot live and work where they are most efficient. No matter how many brilliant new ideas a society has, they are functionally useless if they cannot be built when they are needed. It is up to governments to stop the nimbys. Technology is not from one place or another. The same is true of energy, where permits and infrastructure are what keep costs sky-high.
The AI ​​economy may even become less productive. Let’s examine some recent technologies. Smartphones allow for instant communication, but they can also be distracting. With email you are connected 24 hours a day, seven days a week, which can make it difficult to concentrate. According to a 2016 study by researchers at the University of California at Irvine, Microsoft Research, and MIT, “the more time spent on email each day, the lower the perceived productivity.†Some bosses now believe that working from home, once seen as a productivity booster, gives too many people an excuse to waste time.
Generative artificial intelligence could act as a drag on productivity. What would happen, for example, if it were capable of creating entertainment perfectly adapted to all our desires? Also, few have thought about the consequences of a system that generates large amounts of text instantly. GPT-4 is a godsend for a nimby faced with a planning request. In five minutes he produces a well-written 1,000-page objection. And someone will have to answer to it. Spam emails will be harder to detect. Fraud cases could skyrocket. Banks will have to spend more on preventing attacks and compensating those affected.
In a world full of artificial intelligence, lawyers will multiply. “In the 1970s, you could write a multi-million dollar deal in 15 pages because retyping something was such a drag,†says Preston Byrne of the Brown Rudnick law firm. “Artificial intelligence will allow us to cover the 1,000 most likely edge cases in a first draft, and then the parties will discuss it for weeks.†In the United States, there is a golden rule according to which there is no point in suing for damages if you do not expect to obtain compensation of at least $250,000, which is the amount that you have to spend to get to court. Now the costs of the lawsuit could be reduced to almost zero. Instead, teachers and editors will have to check that everything they read has not been composed by artificial intelligence. OpenAI has released a program that allows you to do this. In this way, it offers the world a solution to a problem that its technology has created.
Artificial intelligence has the potential to change the world in ways that are impossible to imagine today. That’s not the same as turning the economy upside down anyway. As Fogel points out in his study: “The reasoning offered is not intended to refute the view that the railroad played a decisive role in American development during the nineteenth century, but rather to show that the empirical basis on which that view rests is far from as solid as is usually supposed. Sometime in the mid-21st century, a future Nobel Prize winner looking at generative artificial intelligence might come to the same conclusion.
© 2023 The Economist Newspaper Limited. All rights reserved. Translation: Juan Gabriel López Guix
