by Kai-Fu Lee
For the “Human Veneer” quadrant, much of the computational or physical work can already be done by machines, but the key social interactive element makes them difficult to automate en masse. The name of the quadrant derives from the most likely route to automation: while the behind-the-scenes optimization work is overtaken by machines, human workers will act as the social interface for customers, leading to a symbiotic relationship between human and machine. Jobs in this category could include bartender, schoolteacher, and even medical caregiver. How quickly and what percentage of these jobs disappear depends on how flexible companies are in restructuring the tasks done by their employees, and how open customers are to interacting with computers.
The “Slow Creep” category (plumber, construction worker, entry-level graphic designer) doesn’t rely on human beings’ social skills but instead on manual dexterity, creativity, or ability to adapt to unstructured environments. These remain substantial hurdles for AI, but ones that the technology will slowly chip away at in the coming years. The pace of job elimination in this quadrant depends less on process innovation at companies and more on the actual expansion in AI capabilities. But at the far right end of the “Slow Creep” are good opportunities for the creative professionals (such as scientists and aerospace engineers) to use AI tools to accelerate their progress.
These graphs give us a basic heuristic for understanding what kinds of jobs are at risk, but what does this mean for total employment on an economy-wide level? For that, we must look to the economists.
WHAT THE STUDIES SAY
Predicting the scale of AI-induced job losses has become a cottage industry for economists and consulting firms the world over. Depending on which model one uses, estimates range from terrifying to totally not a problem. Here I give a brief overview of the literature and the methods, highlighting the studies that have shaped the debate. Few good studies have been done for the Chinese market, so I largely stick to studies estimating automation potential in the United States and then extrapolate those results to China.
A pair of researchers at Oxford University kicked things off in 2013 with a paper making a dire prediction: 47 percent of U.S. jobs could be automated within the next decade or two. The paper’s authors, Carl Benedikt Frey and Michael A. Osborne, began by asking machine-learning experts to evaluate the likelihood that seventy occupations could be automated in the coming years. Combining that data with a list of the main “engineering bottlenecks” in machine learning (similar to the characteristics denoting the “Safe Zone” in the graphs on pages 155 and 156), Frey and Osborne used a probability model to project how susceptible an additional 632 occupations are to automation.
The result—that nearly half of U.S. jobs were at “high risk” in the coming decades—caused quite a stir. Frey and Osborn were careful to note the many caveats to their conclusion. Most importantly, it was an estimate of what jobs it would be technically possible to do with machines, not actual job losses or resulting unemployment levels. But the ensuing flurry of press coverage largely glossed over these important details, instead warning readers that half of all workers would soon be out of a job.
Other economists struck back. In 2016, a trio of researchers at the Organization for Economic Cooperation and Development (OECD) used an alternate model to produce an estimate that seemed to directly contradict the Oxford study: just 9 percent of jobs in the United States were at high risk of automation.
Why the huge gap? The OECD researchers took issue with Osborne and Frey’s “occupation-based” approach. While the Oxford researchers asked machine-learning experts to judge the automatability of an occupation, the OECD team pointed out that it’s not entire occupations that will be automated but rather specific tasks within those occupations. The OECD team argued that this focus on occupations overlooks the many different tasks an employee performs that an algorithm cannot: working with colleagues in groups, dealing with customers face-to-face, and so on.
The OECD team instead proposed a task-based approach, breaking down each job into its many component activities and looking at how many of those could be automated. In this model, a tax preparer is not merely categorized as one occupation but rather as a series of tasks that are automatable (reviewing income documents, calculating maximum deductions, reviewing forms for inconsistencies, etc.) and tasks that are not automatable (meeting with new clients, explaining decisions to those clients, etc.). The OECD team then ran a probability model to find what percentage of jobs were at “high risk” (i.e., at least 70 percent of the tasks associated with the job could be automated). As noted, they found that in the United States only 9 percent of workers fell in the high-risk category. Applying that same model on twenty other OECD countries, the authors found that the percentage of high-risk jobs ranged from just 6 percent in Korea to 12 percent in Austria. Don’t worry, the study seemed to say, reports of the death of work have been greatly exaggerated.
Unsurprisingly, that didn’t settle the debate. The OECD’s task-based approach came to hold sway among researchers, but not all of them agreed with the report’s sanguine conclusions. In early 2017, researchers at PwC used the task-based approach to produce their own estimate, finding instead that 38 percent of jobs in the United States were at high risk of automation by the early 2030s. It was a striking divergence from the OECD’s 9 percent, one that stemmed simply from using a slightly different algorithm in the calculations. Like the previous studies, the PwC authors are quick to note that this is merely an estimate of what jobs could be done by machines, and that actual job losses will be mitigated by regulatory, legal, and social dynamics.
After these wildly diverging estimates, researchers at the McKinsey Global Institute landed somewhere in the middle. I assisted the institute in its research related to China and coauthored a report with it on the Chinese digital landscape. Using the popular task-based approach, the McKinsey team estimated that around 50 percent of work tasks around the world are already automatable. For China, that number was pegged at 51.2 percent, with the United States coming in slightly lower, at 45.8 percent. But when it came to actual job displacement, the McKinsey researchers were less pessimistic. If there is rapid adoption of automation techniques (a scenario most comparable to the above estimates), 30 percent of work activities around the world could be automated by 2030, but only 14 percent of workers would need to change occupations.
So where does this survey of the literature leave us? Experts continue to be all over the map, with estimates of automation potential in the United States ranging from just 9 percent to 47 percent. Even if we stick to only the task-based approach, we still have a spread of 9 to 38 percent, a divide that could mean the difference between broad-based prosperity and an outright jobs crisis. That spread of estimates shouldn’t cause us to throw up our hands in confusion. Instead, it should spur us to think critically about what these studies can teach us—and what they may have missed.
WHAT THE STUDIES MISSED
While I respect the expertise of the economists who pieced together the above estimates, I also respectfully disagree with the low-end estimates of the OECD. That difference is rooted in two disagreements: one in terms of the inputs of their equations, and one major difference in the way I envision AI disrupting labor markets. The quibble causes me to go with the higher-end estimates of PwC, and the difference in vision leads me to raise that number higher still.
My disagreement on inputs stems from the way the studies estimated the technical capabilities of machines in the years ahead. The 2013 Oxford study asked a group of machine-learning experts to predict whether seventy occupations would likely be automated in the coming two decades, using those assessments to project automatability more broadly. And though the OECD and PwC studies differed in how they divided up occupations and tasks, they basically stuck with the 2013 estimates of future capabilities.
Those estimates probably constituted the best guess of experts at the time, but significant advances in the accuracy and power of machine learn
ing over the past five years have already moved the goalposts. Experts back then may have been able to project some of the improvements that were on the horizon. But few, if any, experts predicted that deep learning was going to get this good, this fast. Those unexpected improvements are expanding the realm of the possible when it comes to real-world uses and thus job disruptions.
One of the clearest examples of these accelerating improvements is the ImageNet competition. In the competition, algorithms submitted by different teams are tasked with identifying thousands of different objects within millions of different images, such as birds, baseballs, screwdrivers, and mosques. It has quickly emerged as one of the most respected image-recognition contests and a clear benchmark for AI’s progress in computer vision.
When the Oxford machine-learning experts made their estimates of technical capabilities in early 2013, the most recent ImageNet competition of 2012 had been the coming-out party for deep learning. Geoffrey Hinton’s team used those techniques to achieve a record-setting error rate of around 16 percent, a large leap forward in a competition where no team had ever gotten below 25 percent.
That was enough to wake up much of the AI community to this thing called deep learning, but it was just a taste of what was to come. By 2017, almost every team had driven error rates below 5 percent—approximately the accuracy of humans performing the same task—with the average algorithm of that year making only one-third of the mistakes of the top algorithm of 2012. In the years since the Oxford experts made their predictions, computer vision has now surpassed human capabilities and dramatically expanded real-world use-cases for the technology.
Those amped-up capabilities extend far beyond computer vision. New algorithms constantly set and surpass records in fields like speech recognition, machine reading, and machine translation. While these strengthened capabilities don’t constitute fundamental breakthroughs in AI, they do open the eyes and spark the imaginations of entrepreneurs. Taken together, these technical advances and emerging uses cause me to land on the higher end of task-based estimates, namely, PwC’s prediction that 38 percent of U.S. jobs will be at high risk of automatability by the early 2030s.
TWO KINDS OF JOB LOSS: ONE-TO-ONE REPLACEMENTS AND GROUND-UP DISRUPTIONS
But beyond that disagreement over methodology, I believe using only the task-based approach misses an entirely separate category of potential job losses: industry-wide disruptions due to new AI-empowered business models. Separate from the occupation- or task-based approach, I’ll call this the industry-based approach.
Part of this difference in vision can be attributed to professional background. Many of the preceding studies were done by economists, whereas I am a technologist and early-stage investor. In predicting what jobs were at risk of automation, economists looked at what tasks a person completed while going about their job and asked whether a machine would be able to complete those same tasks. In other words, the task-based approach asked how possible it was to do a one-to-one replacement of a machine for a human worker.
My background trains me to approach the problem differently. Early in my career, I worked on turning cutting-edge AI technologies into useful products, and as a venture capitalist I fund and help build new startups. That work helps me see AI as forming two distinct threats to jobs: one-to-one replacements and ground-up disruptions.
Many of the AI companies I’ve invested in are looking to build a single AI-driven product that can replace a specific kind of worker—for instance, a robot that can do the lifting and carrying of a warehouse employee or an autonomous-vehicle algorithm that can complete the core tasks of a taxi driver. If successful, these companies will end up selling their products to companies, many of whom may lay off redundant workers as a result. These types of one-to-one replacements are exactly the job losses captured by economists using the task-based approach, and I take PwC’s 38 percent estimate as a reasonable guess for this category.
But then there exists a completely different breed of AI startups: those that reimagine an industry from the ground up. These companies don’t look to replace one human worker with one tailor-made robot that can handle the same tasks; rather, they look for new ways to satisfy the fundamental human need driving the industry.
Startups like Smart Finance (the AI-driven lender that employs no human loan officers), the employee-free F5 Future Store (a Chinese startup that creates a shopping experience comparable to the Amazon Go supermarket), or Toutiao (the algorithmic news app that employs no editors) are prime examples of these types of companies. Algorithms aren’t displacing human workers at these companies, simply because the humans were never there to begin with. But as the lower costs and superior services of these companies drive gains to market share, they will apply pressure to their employee-heavy rivals. Those companies will be forced to adapt from the ground up—restructuring their workflows to leverage AI and reduce employees—or risk going out of business. Either way, the end result is the same: there will be fewer workers.
This type of AI-induced job loss is largely missing from the task-based estimates of the economists. If one applied the task-based approach to measuring the automatability of an editor at a news app, you would find dozens of tasks that can’t be performed by machines. They can’t read and understand news and feature articles, subjectively assess appropriateness for a particular app’s audience, or communicate with reporters and other editors. But when Toutiao’s founders built the app, they didn’t look for an algorithm that could perform all of the above tasks. Instead, they reimagined how a news app could perform its core function—curate a feed of news stories that users want to read—and then did that by employing an AI algorithm.
I estimate this kind of from-the-ground-up disruption will affect about 10 percent of the workforce in the United States. The hardest hit industries will be those that involve high volumes of routine optimization work paired with external marketing or customer service: fast food, financial services, security, even radiology. These changes will eat away at employment in the “Human Veneer” quadrant of the earlier chart, with companies consolidating customer interaction tasks into a handful of employees, while algorithms do most of the grunt work behind the scenes. The result will be steep—though not total—reductions in jobs in these fields.
THE BOTTOM LINE
Putting together percentages for the two types of automatability—38 percent from one-to-one replacements and about 10 percent from ground-up disruption—we are faced with a monumental challenge. Within ten to twenty years, I estimate we will be technically capable of automating 40 to 50 percent of jobs in the United States. For employees who are not outright replaced, increasing automation of their workload will continue to cut into their value-add for the company, reducing their bargaining power on wages and potentially leading to layoffs in the long term. We’ll see a larger pool of unemployed workers competing for an even smaller pool of jobs, driving down wages and forcing many into part-time or “gig economy” work that lacks benefits.
This—and I cannot stress this enough—does not mean the country will be facing a 40 to 50 percent unemployment rate. Social frictions, regulatory restrictions, and plain old inertia will greatly slow down the actual rate of job losses. Plus, there will also be new jobs created along the way, positions that can offset a portion of these AI-induced losses, something that I explore in coming chapters. These could cut actual AI-induced net unemployment in half, to between 20 and 25 percent, or drive it even lower, down to just 10 to 20 percent.
These estimates are in line with those from the most recent research (as of this writing) that attempted to put a number on actual job losses, a February 2018 study by the consulting firm Bain and Company. Instead of wading into the minutiae of tasks and occupations, the Bain study took a macro-level approach, seeking to understand the interplay of three major forces acting on the global economy: demographics, automation, and inequality. Bain’s analysis produced a startling bottom-line conclusion: by 2030, employers will need 20 to 25 p
ercent fewer employees, a percentage that would equal 30 to 40 million displaced workers in the United States.
Bain acknowledged that some of these workers will be reabsorbed into new professions that barely exist today (such as robot repair technician), but predicted that this reabsorption would fail to make a meaningful dent in the massive and growing trend of displacement. And automation’s impact will be felt far wider than even this 20 to 25 percent of displaced workers. The study calculated that if we include both displacement and wage suppression, a full 80 percent of all workers will be affected.
This would constitute a devastating blow to working families. Worse still, this would not be a temporary shock, like the fleeting brush with 10 percent unemployment that the United States experienced following the 2008 financial crisis. Instead, if left unchecked, it could constitute the new normal: an age of full employment for intelligent machines and enduring stagnation for the average worker.
U.S.-CHINA COMPARISON: MORAVEC’S REVENGE
But what about China? How will its workers fare in this brave new economy? Few good studies have been conducted on the impacts of automation here, but the conventional wisdom holds that Chinese people will be hit much harder, with intelligent robots spelling the end of a golden era for workers in the “factory of the world.” This prediction is based on the makeup of China’s workforce, as well as a gut-level intuition about what kinds of jobs become automated.