The question now is: What can and can’t computers do in the age of AI?
Identifying engineering bottlenecks to automation is obviously not an economics question, so I was fortunate that Michael had been researching this subject for some time. The trouble economists face in researching technological change is that we are bound to be behind the curve (Simon was not just an economist but also a highly regarded computer scientist). I was hardly up to date on everything going on in the labs. But as I was writing economics papers, Michael had been developing the algorithms that have expanded the set of tasks computers are now able to perform.
In the spirit of Simon, we set out to determine in which domains humans still retain the comparative advantage. Rather than asking unanswerable questions about the prospects of superintelligence or trying to predict the great inventions of the future, we looked at technologies on the horizon. In the words of Thomas Malthus, writing at the onset of the Industrial Revolution, “many discoveries have already taken place in the world that were totally unforeseen and unexpected.… But if a person had predicted these discoveries without being guided by any analogies or indications from past facts, he would deserve the name of seer or prophet, but not of philosopher.”48 Many of the technologies discussed here are still prototypes, but their arrival in the marketplace is not unforeseeable, and while they are still imperfect, every technological revolution began with imperfect technology. The early steam engines were used only to drain mines, and they did not even do that particularly well. Yet Thomas Savory, Thomas Newcomen, and James Watt, all realized that the steam engine was a GPT, and they conceived many applications for it. As noted above, AI is another GPT, and it is already being used to perform both mental and manual tasks.
Because its potential applications are so vast, Michael and I began by looking at tasks that computers still perform poorly and where technological leaps have been limited in recent years. For a glimpse of the state of the art in machine social intelligence, for example, consider the Turing test, which captures the ability of an AI algorithm to communicate in a way that is indistinguishable from an actual human. The Loebner Prize is an annual Turing test competition that awards prizes to chat bots that are considered to be the most humanlike. These competitions are straightforward. A human judge holds computer-based textual interactions with both an algorithm and a human at the same time. Based on these conversations, the judge must then try to distinguish between the two. In a paper written in 2013, Michael and I noted: “Sophisticated algorithms have so far failed to convince judges about their human resemblance.”49 Yet a year later the computer program Eugene Goostman managed to convince 33 percent of the judges that it was a person. Some people subsequently argued that we had underestimated the accelerating pace of change. However, such claims exaggerate the capabilities of Eugene Goostman, which simulated a thirteen-year-old boy speaking English as his second language. Even if we assume that algorithms at some point will be able to effectively reproduce human social intelligence in basic texts, many jobs center on personal relationships and complex interpersonal communication. Computer programmers consult with managers or clients to clarify intent, identify problems, and suggest changes. Nurses work with patients, families, or communities to design and implement programs to improve overall health. Fund-raisers identify potential donors and build relationships with them. Family therapists counsel clients on unsatisfactory relationships. Astronomers build research collaborations and present their findings on conferences. These tasks are all way beyond the competence of computers.
Many jobs also require creativity, like the ability to come up with new, unusual, and clever ideas. Survey data show that the work of physicists, art directors, comedians, CEOs, video game designers, and robotics engineers, to name a few, all involve such activities.50 The challenge here, from an automation point of view, is not one of generating novelty but generating novelty that makes sense. For a computer to produce an original piece of music, write a novel, develop a new theory or product, or make a subtle joke, in principle the only things that are required is a database with a richness of experience that is comparable to that of humans and solid methods that allow us to benchmark the algorithm’s subtlety. It is also entirely possible to give an algorithm access to a database of symphonies, label some as bad and others as good, and allow it to generate an original recombination. Algorithms already exist that generate music in many different styles, reminiscent of specific human composers. But people do not just generate ideas on the basis of related existing works. They draw upon experiences from all aspects of life.
As discussed above, many challenges remain when algorithms have to interact with a variety of irregular objects in uncontrolled environments. Some perception tasks, like identifying objects and their properties in a cluttered field of view, have proven hard to overcome. Robots remain unable to match the depth and breadth of human perception, which translates into further difficulties in manipulation. Distinguishing a pot that is dirty and needs to be cleaned from a pot holding a plant is straightforward to any human. Yet robots still struggle to mimic human capabilities in such tasks, making many jobs—like those of janitors and cleaners—exceedingly hard to automate. Though single-purpose robots exist, capable of doing individual tasks like cleaning floors, there is no multipurpose robot that could find and remove rubbish. In controlled environments like factories and warehouses, it is possible to circumvent some engineering bottlenecks through clever task redesign. But a home is a different matter. In addition to the hard perception tasks, like identifying rubbish, there are “yet further problems in designing manipulators that, like human limbs, are soft, have compliant dynamics and provide useful tactile feedback.”51 While there has been recent progress in very simple tasks like spinning an alphabet block, picking up familiar objects with a gripper, and even teaching robots to recognize the best way of picking up any item using AI, the ability of advanced robotics to manipulate various objects is still very limited. Most industrial manipulation makes use of work-arounds to address these challenges.
With these engineering bottlenecks in mind, Michael and I proceeded to explore the automatability of jobs based on twenty thousand unique task descriptions.52 This sort of detailed information comes with one problem: it is a lot of data to process. So instead of examining each individual task, we took a sample of seventy occupations that a group of AI experts deemed either automatable or nonautomatable on the basis of the tasks the occupations entail. This gave us what machine learning researchers call a training data set. While the task descriptions for each occupation were unique, our database also provided a set of common features. Based on these features, our algorithm was able to learn about the characteristics of automatable occupations, allowing it to predict the exposure to automation of another 632 occupations. Thus, our final sample covered 702 occupations, in which 97 percent of the American workforce is employed.
Using AI for our analysis did not just have the benefit of saving time and labor. Our analysis also underlined the fact that algorithms are now infinitely superior to humans in pattern recognition. We were quite convinced that the work of waiters and waitresses would not lend itself to automation, but our algorithm told us that we were wrong. Analyzing the similarities between the jobs of waiters and other jobs in a more comprehensive manner than we possibly could have done, it predicted that waiters are susceptible to automation. Indeed, in the months after our original analysis, we learned that McDonald’s had plans to install self-ordering kiosks. We learned about Chili’s Grill and Bar’s plans to complete the rollout of its tablet ordering system. We learned that Applebee would introduce tablets to eighteen hundred restaurants. And in 2016, a new and almost fully automated restaurant chain called Eatsa opened. Customers order their food at an iPad kiosk. They then wait a few minutes in front of a giant vending machine that churns out freshly prepared quinoa bowls. At the other side of the machine, kitchen staff members cook the food, but Eatsa does not employ any waiters.
&nb
sp; Of course, this does not mean that all waiting jobs will be replaced. In many settings, consumers are likely to prefer human waiters as part of the service experience. What it does tell us is that their jobs are automatable in principle. We shall return to the question of the determinants of technology adoption shortly.
FIGURE 17: Share of Jobs at Risk of Automation by Major Occupational Categories
Source: C. B. Frey and M. A. Osborne, 2017, “The Future of Employment: How Susceptible Are Jobs to Computerisation?,” Technological Forecasting and Social Change 114 (January): 254–80.
Figure 17 plots the exposure of major occupational categories to automation by their employment share. Office and administrative support, production, transportation and logistics, food preparation, and retail jobs loom large in terms of both their exposure to automation and the percentage of Americans they support. Overall, our algorithm predicted that 47 percent of American jobs are susceptible to automation, meaning that they are potentially automatable from a technological point of view, given the latest computer-controlled equipment and sufficient relevant data for the algorithm to draw upon. What most of these jobs have in common is that they are low-income jobs that do not require high levels of education (figure 18).
FIGURE 18: Jobs at Risk of Automation by Income and Educational Attainment
Source: C. B. Frey and M. A. Osborne, 2017, “The Future of Employment: How Susceptible Are Jobs to Computerisation?,” Technological Forecasting and Social Change 114 (January): 254–80.
Note: The figure plots the probability that an occupation is automatable against the median annual income and educational attainment in that occupation. It shows that occupations in which people earn higher wages and have higher educational attainment are less exposed to automation on average.
A number of studies have emerged since we first published our article, reaching somewhat different conclusions. Research by the OECD, for example, estimates that 14 percent of jobs are at risk of being replaced, with another 32 percent being at risk of significant change.53 The OECD mistakenly argues that we overestimated the scope of automation by focusing on occupations rather than tasks. What they miss is that we inferred the automatability of occupations on the basis of the tasks they entail. According to our estimates, medical doctors are not at risk of automation even if algorithms are becoming more pervasive in tasks like medical diagnostics. And journalists are not exposed to automation just because AI algorithms are now capable of churning out shorter news stories. Both journalists and medical doctors are safe from automation, according to our estimates, even though they entail some tasks that lend themselves to automation. So why, then, do the OECD’s estimates diverge so much from ours? One explanation is that they use less-detailed occupational data. Another is that their model performs less well against our training data set.54
However, for all their differences, these studies concur that unskilled jobs are most exposed to automation.55 When President Barack Obama’s Council of Economic Advisers used our estimates to sort by wage levels the occupations most at risk of being automated, they found that 83 percent of workers in occupations that paid less than $20 an hour were at high risk of being replaced, while the corresponding figure for workers in occupations that paid more than $40 per hour was only 4 percent.56 What this shows is that the labor market prospects of the unskilled will likely continue to deteriorate, unless other forces counteract that trend. We saw in chapter 9 that the first wave of automation of routine work pushed many Americans out of decent middle-class jobs and into low-paying service jobs. Many of these low-skilled jobs are now threatened by automation, too. If anything, the next wave can be expected to put more downward pressure on the wages of the middle class, many members of which are already competing for low-income jobs. In the words of Harvard University’s Jason Furman, who chaired Obama’s Council of Economic Advisers, “We are already seeing some of this play out—for example, when we go shopping and take our groceries to a kiosk instead of a cashier, or when we call a customer service help line and interact with an automated customer service representative.”57
Thus, a widespread misconception is that automation is coming for the jobs of the skilled. In the best-selling Rise of the Robots, Martin Ford declared that “employment for many skilled professionals—including lawyers, journalists, scientists, and pharmacists—is already being significantly eroded by advancing information technology [so that] acquiring more education and skills will not necessarily offer effective protection against job automation in the future.”58 Though many of the jobs Ford highlights surely involve some tasks that can be automated away, they also involve many more tasks that cannot. For example, when Dana Remus and Frank Levy recently analyzed lawyers’ billing records, they found that if AI and several related applications were all adopted immediately—which seems highly unlikely—this would substitute for roughly 13 percent of their time. Most of their time is spent performing tasks like legal writing, investigating facts, negotiating, appearing in court, and advising clients. As Remus and Levy explain, lawyers’ work requires more than prediction: “It requires a lawyer to understand a client’s situation, goals, and interests; to think creatively about how best to serve those interests pursuant to law; and sometimes, to push back against a client’s proposed course of action and counsel compliance. These are things that frequently require human interaction and emotional intelligence and cannot, at least for the time being, be automated.”59 Reassuringly, our algorithm also predicted that lawyers are at low risk of automation.
Amara’s Law
Though the scope of automation is significant, its pace is a different matter. Like Simon’s forecast, our predictions were based on merely observing what computers can do. And like Simon, we did not try to predict the pace of change, which depends on many unpredictable factors beyond the technology itself.60 But we surely didn’t expect 47 percent of jobs to be automated anytime soon. We highlighted a wide range of factors that are likely to shape the pace of automation. The bottom line is that all the prototype technologies discussed here will not arrive at the same time. And their adoption will not be frictionless, either. Regulation, consumer preferences, and worker opposition, among many other variables, will shape the speed of adoption. Thus, inflated expectations have typically been followed by disillusionment. As Roy Amara famously observed, “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.” Indeed, Amara’s Law has been a good guide to the trajectories of technological progress in the past.
Seen through the lens of history, the scope of automation this time around is probably not as staggering as has sometimes been suggested. In 1870, around 46 percent of the American workforce was still employed in agriculture, while today the agriculture sector absorbs about 1 percent of the labor force (table 1).61 Tractors played a key role in reducing labor requirements on the farms (see chapters 6 and 8). But while one might have inferred that many farm jobs were at risk of replacement when the gasoline-powered tractor arrived, the speed of adoption would have been much harder to predict.
The hurdles to tractor adoption were many. First, increasingly complex machinery required more skilled operators. Early on, farmers typically waited to purchase tractors, wanting to see how long it took for laborers on other farms to acquire the mechanical skill required. As an article in the New York Times reported in 1918, “A tractor is a too good machine to put in the hands of a poor operator.… Where to get first-class tractor operators is often more of a puzzle to the buyer than how to get the machine.”62 In the same year, the New York State College of Agriculture announced a three-week course for tractor and truck operators, to bridge the skills gap and accelerate adoption. Second, the adoption of tractors—like other GPTs—moved at different speeds across applications: “The earliest models were suitable only for tillage and harvesting small grains, and only in the late 1920s did the technology begin to generalize for use with row crops such as corn, cotton, and vegetables.”
63 Some use cases emerged only in the later stages of mechanical development.
Third, even as tractors became more pervasive, the abundance of cheap labor in the countryside meant that the mechanization of farming did not make economic sense for a long time. As the Second Industrial Revolution produced an ever-growing number of well-paying industrial jobs, however, many Americans left the farms for the cities, which increased incentives to mechanize. But even so, tractors were still not economically viable in many settings. They were mainly used on large farms that relied on wage labor. Many low-income farmers were highly risk averse and preferred to continue to rely on horses rather than investing in expensive tractors—even though this meant that they had to set aside acreage to grow feed. If a tractor could not be bought outright, loan payments were still significant enough to deter adoption. In 1921, the New York Times pointed out that there were still seventeen million horses on American farms but only 246,139 tractors and expressed concern that adoption was lagging. A push, the article argued, was required to increase productivity in agriculture.64 And a push followed a decade later. One reason that adoption finally accelerated in the 1930s, despite a decade of the Great Depression, were New Deal programs like the Commodity Credit Corporation and the Farm Credit Administration, which reduced the price risk, lowered interest rates, and made cash available to farmers.65
* * *
Amara’s Law applied to the computer revolution, too. Despite widespread automation anxiety in the 1950s and 1960s (chapter 7), computers were too bulky and expensive to find widespread use before the 1980s (see chapter 9). While many businesses marveled at the capabilities of computers, few opened their wallets. As risk-adverse farmers were reluctant to adopt expensive tractors, businesses deemed the cost of computers too high to bear. And they were right in thinking so. When computerization finally took off, unforeseen glitches emerged. In 1987, when Robert Solow puzzled that “we can see the computer age everywhere but in the productivity statistics,” an article in the Wall Street Journal reported: “Companies are automating in smaller doses now, a strategy that allows bugs to be worked out before huge investments are made.”66 As the director of engineering at AT&T explained, “If you make 30 million boxes of Wheaties a year, you can use automation without many problems, but if you’re in a competitive market where the product is changing and its life cycle is short, you better be damned careful.”67
The Technology Trap Page 36