The graph below, based on data from the Federal Reserve Bank of St. Louis, shows manufacturing employment in the United States as a fraction of all employment. As you can see, the line heads downward in an almost perfectly straight line beginning in the mid-1950s. Notice that the line doesn’t become steeper as globalization takes hold after the passage of NAFTA in 1994 or the rise of China over the past decade or so. The line just slopes consistently downward.
This is primarily the result of technology, and in particular, automation. Manufacturing in the U.S. has become dramatically more productive and requires fewer workers. If we were to graph manufacturing output (rather than jobs), the line would slope upward, not downward. The value of U.S. manufacturing production is now far greater than it was in industrial era of the 1950s, even after adjusting for inflation. We just make all that stuff with a lot fewer people.
One of the most interesting things about the graph above is that, if technology is the primary driver, then employment in China must inevitably follow the same path. In fact, there are good reasons to believe that manufacturing employment’s downward slope will be significantly steeper for China. The U.S. had to invent the technology to make manufacturing more productive, while in many cases China only needs to import it from more developed nations. It is also true that China is beginning its journey at a time when information technology (which is the primary enabler of automation) is many orders of magnitude more advanced than in the 1950s when U.S. manufacturing employment was at its peak. (See this recent article on skilled robots from the New York Times).
In the U.S. (as well as in other advanced countries), workers shifted out of manufacturing and into the service sector — which now accounts for the vast majority of jobs. Will China be able to pull off the same transition?
The U.S. had the luxury of building a strong middle class during an earlier time. Technology was advancing consistently and increasing productivity, but it was not so advanced as to create a mismatch between the type of available jobs and the skills of workers. Unionization was strong in the private sector and helped ensure that the lion’s share of productivity increases ended up in workers’ (rather that corporate owners’) pockets. Those workers, in turn, became the broad-based consumer class that purchased the output from all those factories and kept the overall economy humming.
The situation in China is quite different. Consumer spending accounts for only about a third of China’s GDP (as opposed to 60% or more in nearly all developed countries). While China has built a significant middle class in absolute terms, it remains small as a percentage of the country’s huge population.
Workers enjoy few of the rights and protections that characterized the U.S. workforce of the 1950s. As I wrote in my book, The Lights in the Tunnel:
The [Chinese] government actively enforces discrimination that tends to drive wages even lower. Much of the work in China’s factories is performed by migrant workers who officially live in the countryside but are allowed to come to cities or industrial regions to work. These workers typically live in factory dormitories and do not have the right to bring their families to the cities or to genuinely assimilate into an urban middle class. Wages for these workers are far lower than for urban dwellers, and the money that they do earn is for the most part either saved or sent home to help support their families. These workers are not in a position to become major drivers of local consumption any time soon.
According to the New York Times, those worker dormitories apparently play an important role in Apple’s (or Foxconn’s) ability to bring production online at any time of the night or day:
A foreman immediately roused 8,000 workers inside the company’s dormitories, according to the executive. Each employee was given a biscuit and a cup of tea, guided to a workstation and within half an hour started a 12-hour shift fitting glass screens into beveled frames. Within 96 hours, the plant was producing over 10,000 iPhones a day.
Even that level of worker availability and efficiency isn’t enough for Foxconn, which recently announced the introduction of huge numbers of robots. That may be a great way to drive production, but it’s hard to see how China will succeed in dramatically shifting its economy toward domestic consumption.
And that has to happen before a shift to a service economy can take place. As consumers become more wealthy they begin to spend a larger fraction of their incomes on services — things like banking, insurance, healthcare, education, entertainment and travel — and that in turn drives service sector employment. At least that has been the path followed in other developed countries.
In the absence of consumer spending, China’s economy remains highly dependent on manufacturing exports and, especially, on fixed investment. An astonishing 50% of China’s GDP is driven by investment in things like factories, housing and infrastructure (the U.S. figure is around 15%). The problem is that all that investment has to ultimately pay for itself, and that happens via consumption. Once a factory is built it has to then produce something that gets sold at a profit. Homes, retail buildings and apartment complexes likewise have to be sold or rented out. Obviously, no economy can indefinitely invest anything like 50% of its output without eventually finding a way to get a positive return on that investment.
Achieving that return requires consumers — either at home or abroad. China continues to rely heavily on consumers in the U.S. and Europe, but that’s unlikely to be a sustainable formula for growth. The debt crisis and the resulting austerity is cutting into economic growth and consumer spending in both Europe and the U.S.
As manufacturing automation increases (perhaps dramatically) in China, in the U. S. and other developed countries the most disruptive impact from technology will be in the service sector — where millions of white collar jobs and service jobs in retail, distribution, food service and other areas may ultimately be at risk. After all, if robots can build an iPhone, then its a good bet that they will also someday be able to build a hamburger or mix a latte. The result may be continuing high unemployment, stagnant wages and tepid consumer spending throughout much of the developed world.
The real problem China faces is that it is late to the party. Just as it reaches its manufacturing employment zenith, it faces a potentially disruptive impact from automation technology. And that will happen roughly in parallel with similar transitions in the service sectors of the countries that currently consume much of its output. In the face of that, can China succeed in re-balancing its economy toward consumption, increasing personal incomes, and building a vibrant service sector to keep its population employed?
PBS News Hour recently had a special on the main topic I’ve been focusing on here: unemployment and inequality caused by technology, and in particular, automation. You can watch the video below.
At around 05:40, Ray Kurzweil makes a brief appearance. He is asked about the possibility of a “digital divide” — meaning that only a small percentage of the population is able to take advantage of new technologies, even as traditional employment opportunities are destroyed. Kurzweil seems to argue that we won’t have a problem because these new technologies will be affordable and widely available (he gives the example of cell phones). A little later in the video, Peter Diamandis, the chairman of Singularity University, makes essentially the same point.
These views strike me as both unrealistic and elitist. There is little evidence to suggest that most average people are going to be able to parlay access to a cell phone, social media, or other personal technologies into a livable income. Even among the minority of people who actually have the necessary skills and training, there is a strong element of luck associated with the success of any entrepreneurial activity. Most new businesses of any type fail. Assuming that a huge percentage (perhaps most) of the population will someday generate a meaningful income by independently leveraging technology is really quite a stretch.
A second problem with techno-optimists like Kurzweil and Diamandis is their near exclusive focus on the cost side of technology. Many technologists believe that advancing technology and increased automation are likely to drive down costs and possibly make most products and services far more affordable. At the extreme, some techno-optimists believe in the promise of a “post scarcity” economy. Even if we go along with that — and there are certainly powerful opposing arguments based on energy and resource depletion and environmental degradation — simply making “stuff” cheaper is not an adequate solution.
Imagine for a moment that you were living in the year 1900. Suppose you could look through a time portal and see the world of 2012. You might well suppose that a “post scarcity” world had already been realized given the far higher living standards that average people now enjoy. On the other hand, if you got a look at 2012 prices (as opposed to what you were used to in 1900) you certainly wouldn’t feel that things had become more affordable!
The reality, of course, is that prices have increased dramatically in nominal terms since 1900 — but average incomes have increased even more. The average U.S. worker in 1900 earned just $438 per year. Over the past 112 years, incomes have increased dramatically in real terms (after adjusting for inflation), leaving nearly everyone better off, even as prices have increased.
The problem is that if, rather than a period of 112 years, we look at just the last 30 years — say since the mid 1980s — the story is very different. Incomes (wages) for most average workers have been completely stagnant in real terms; after adjusting for inflation, most workers have made little or any progress. And for a number of big ticket items — like health care, housing and education — the situation has actually worsened significantly for most Americans.
So will making all kinds of stuff cheaper, even as incomes continue to stagnate and even fall, solve our problems? No, it will not. If we actually had a situation where prices for nearly everything fell while wages likewise fell and unemployment increased, that would be deflation. You won’t find many economists who would advocate long-term deflation as a good strategy for the future.
Deflation destroys the incentive to invest in the future, and if prolonged, would likely slow the pace of innovation. The problem with deflation is that while incomes, prices and asset values may well fall, debts do not deflate. The result would be widespread insolvency, potentially catastrophic financial crises, and lower living standards for virtually everyone.
The true challenge we face in the future is really about incomes. As technology and globalization advance, how do we get incomes for the majority of the population to continue increasing in real terms? This has been the historical path to prosperity, and we have to figure out how to maintain that trend going forward. One of the main ideas I focus on in my book The Lights in the Tunnel is that incomes power consumers — and consumers ultimately power the economy.
If we can’t find a way to maintain, and even increase, real incomes for the majority of our population, broad-based prosperity will become increasingly elusive.
Ten Jobs that Won’t Be Taken by Robots - The Fiscal Times
Does Facebook Create Jobs? – Pittsburgh Tribune-Review
Is your Job Robot-Proof? – Forbes
Russian Investor Sets Up Robotics VC Fund – The Atlantic
Robot mimics infants’ word learning - Los Angeles Times
Robot Soccer Leads to Innovation – Slate
Robots Learn to Work with Humans – DiscoveryNews
“The Singularity is Near” movie (trailer):
Some recent links showing the on-going march toward job automation:
Automating Legal Work - New York Magazine.
This is more on the the use of e-discovery software to process documents. Automation is likely to hit especially hard at the entry level (and more routine) jobs often taken by recent graduates in a variety of skilled occupations, including law and journalism:
Rapidly Improving Manufacturing Robots – Singularity Hub.
Military Robots – GMA News
Construction Robots – Construction Digital
Productivity Increases are Going to Capital — Not Labor – Paul Krugman
Falling Labor Force Participation Rate – Conversable Economist – As this post points out, demographics (aging workforce) and cyclical factors explain only part of this…
Foxconn, of course, is infamous for the number of its workers who committed suicide. Amazon has had issues of its own. At its Allentown, PA, warehouse, employees were repeatedly overwhelmed by heat and had to seek medical attention. A recent article in Mother Jones tells the story of what it’s like to work in one of these warehouses (the article does not identify the company).
Automation is not just about increasing efficiency. There’s some evidence to suggest that workers are simply being driven beyond their limits. As production speeds continue to increase, there has to come a point where the only option is to get the humans out of the loop. In many industries, automation may penetrate more rapidly than we expect simply because a threshold is reached where people can no longer keep up.
Anyone who is interested in how manufacturing jobs are evolving (and disappearing) should be sure to read Adam Davidson’s excellent article in the current issue of The Atlantic: “Making it in America.” The article is based on interviews with workers and executives at Standard Motor Products, a manufacturer and distributor of auto parts with a factory in Greenville, South Carolina.
Here’s a a quote from the article, focusing on the future prospects for an unskilled worker named Maddie:
Tony [the factory manager] points out that Maddie has a job for two reasons. First, when it comes to making fuel injectors, the company saves money and minimizes product damage by having both the precision and non-precision work done in the same place. Even if Mexican or Chinese workers could do Maddie’s job more cheaply, shipping fragile, half-finished parts to another country for processing would make no sense. Second, Maddie is cheaper than a machine. It would be easy to buy a robotic arm that could take injector bodies and caps from a tray and place them precisely in a laser welder. Yet Standard would have to invest about $100,000 on the arm and a conveyance machine to bring parts to the welder and send them on to the next station. As is common in factories, Standard invests only in machinery that will earn back its cost within two years. For Tony, it’s simple: Maddie makes less in two years than the machine would cost, so her job is safe—for now. If the robotic machines become a little cheaper, or if demand for fuel injectors goes up and Standard starts running three shifts, then investing in those robots might make sense.
“What worries people in factories is electronics, robots,” she tells me. “If you don’t know jack about computers and electronics, then you don’t have anything in this life anymore. One day, they’re not going to need people; the machines will take over. People like me, we’re not going to be around forever.”
As the article makes clear, Maddie has a job largely because she is still cheaper than installing automation equipment — assuming a two-year payback period for the equipment. But now consider that just last year the Boston Globe reported that start-up company Heartland Robotics expects to introduce a manufacturing robot that will sell for around $5,000. The cost of automating jobs like Maddie’s seems likely to fall quite dramatically in the relatively near future. At a cost of $5,000 — or even $10,000 — a robot could easily pay for itself within a matter of months.
Now here’s another quote from the article focusing on a skilled operator named Luke. Luke runs computer-controlled “Gildemeister” machines, which cut precision parts used in fuel injectors:
After six semesters studying machine tooling, including endless hours cutting metal in the school workshop, Luke, like almost everyone who graduates, got a job at a nearby factory, where he ran machines similar to the Gildemeisters. When Luke got hired at Standard, he had two years of technical schoolwork and five years of on-the-job experience, and it took one more month of training before he could be trusted alone with the Gildemeisters. All of which is to say that running an advanced, computer-controlled machine is extremely hard. Luke now works the weekend night shift, 6 p.m. to 6 a.m., Friday, Saturday, and Sunday.
When things are going well, the Gildemeisters largely run themselves, but things don’t always go well. Every five minutes or so, Luke takes a finished part to the testing station—a small table with a dozen sets of calipers and other precision testing tools—to make sure the machine is cutting “on spec,” or matching the requirements of the run. Standard’s rules call for a random part check at least once an hour. “I don’t wait the whole hour before I check another part,” Luke says. “That’s stupid. You could be running scrap for the whole hour.”
The conventional wisdom, of course, is that while Maddie’s job may well be in danger at some point, Luke has little to worry about. Demand for skilled machine operators is strong and is likely to increase in the future. But is that really the way things will play out?
Any time workers are highly paid and in short supply, there is a clear incentive for innovation that will either eliminate those workers or “dumb down” the job so it can be done by less skilled people. Currently, Luke spends a lot of effort testing parts to make sure they remain in spec, and then recalibrating the machine as necessary. Is it possible that machines like the Gildemeisters will someday be able to automatically test parts and then make the required adjustments? Perhaps it might be possible to use precise computer imaging technology to analyze parts and then auto-adjust the machine to produce consistent results. Since this hasn’t happened yet, we can assume the technology probably isn’t there. But what about five or ten years from now?
In a previous post, I wrote about the new “cloud robotics” strategy being pursued by Google, Willow Garage and other companies. Cloud robotics involves migrating much of the software that controls robots or other machines into centralized servers. Could something similar be done with computer-controlled machine tools — so that perhaps dozens of machines are controlled by advanced software that incorporates artificial intelligence, eliminating the need for individual operators?
Yet another possibility involves offshoring. Jobs in areas like customer service can, of course, easily be outsourced to low-wage countries. Could skilled machine operator jobs also be offshored? If the parts are moved around robotically so that the entire job consists of analysis, programming and control, then those tasks could potentially be done from anywhere. China currently has a huge surplus of college graduates — a high percentage of whom have studied science and engineering — and for jobs of this type, English language skills wouldn’t necessarily be critical. Moving an entire factory to China results in major transportation costs and logistics issues. Moving specific jobs offshore electronically, as is currently done in the service sector, might someday prove more cost effective.
In other words, our conventional assumptions about the jobs of the future are not necessarily all that reliable. While Maddie’s job is certainly at risk, even Luke may not turn out to have the level of job security we expect over the longer run.
A final quote is from a discussion with the CEO of the company:
To keep the business of the giant auto-parts retailers, Standard has to constantly lower costs while maintaining quality. High quality is impossible without good raw materials, which Standard has to buy at market rates. The massive global conglomerates, like Bosch, might be able to command discounts when buying, say, specially formulated metals; but Standard has to pay the prevailing price, and for years now, that price has been rising. That places an even higher imperative on reducing the cost of labor. If Standard paid unskilled workers like Maddie more or hired more of them, Larry says, the company would have to charge its customers more or accept lower profits. Either way, Standard would collapse fairly soon.
The main point here is that when a company is squeezed between rising costs for commodity inputs and an inability to raise its own prices, the only significant variable it really has to work with is labor. This dynamic is not limited to manufacturing companies like Standard Parts; previously, I made a similar point about the fast food industry.
Fast Food (or beverage) workers are classified as service workers by the government, but from a technical point of view, fast food is really a kind of just-in-time manufacturing. As automation technology gets better and cheaper, and as the price of food commodities continues to rise in response to ever increasing global demand, it seems likely that the rate of job creation in fast food as well as in a variety of other service areas may be in danger of falling short of expectations.
When it comes to the jobs of the future, beware the conventional wisdom. A great many widely-held assumptions seem to be based primarily on simply projecting the continuation of existing trends, rather than on any meaningful analysis of what the industries of the future might really look like.
At the 2011 Google I/O developer’s conference, Google announced a new initiative called “cloud robotics” in conjunction with robot manufacturer Willow Garage. Willow Garage and a variety of other contributors have developed an open source (free) operating system for robots, with the unsurprising name “ROS” — or Robot Operating System. ROS is being positioned as the MS-DOS (or MS Windows) of robotics.
With ROS and a package called “rosjava“, software developers will be able to write code in the Java programming language and control robots in a standardized way — much in the same way that programmers writing applications for Windows or the Mac can access and control computer hardware.
Google’s approach also offers compatibility with Android. Robots will be able to take advantage of the “cloud-based” (in other words, online) features used in Android phones, as well as new cloud-based capabilities specifically for robots. In essence this means that much of the intelligence that powers the robots of the future may reside on huge server farms, rather than in the robot itself. While that may sound a little “Skynet-esque,” it’s a strategy that could offer huge benefits for building advanced robots.
One of the most important cloud-based robotic capabilities is certain to be object recognition. In my book, The Lights in the Tunnel, I have a section where I talk about the difficulty of building a general-purpose housekeeping robot largely because of the object recognition challenge:
A housekeeping robot would need to be able to recognize hundreds or even thousands of objects that belong in the average home and know where they belong. In addition, it would need to figure out what to do with an almost infinite variety of new objects that might be brought in from outside.
Designing computer software capable of recognizing objects in a very complex and variable field of view and then controlling a robot arm to correctly manipulate those objects is extraordinarily difficult. The task is made even more challenging by the fact that the objects could be in many possible orientations or configurations. Consider the simple case of a pair of sunglasses sitting on a table. The sunglasses might be closed with the lenses facing down, or with the lenses up. Or perhaps the glasses are open with the lenses oriented vertically. Or maybe one side of the glasses is open and the other closed. And, of course, the glasses could be rotated in any direction. And perhaps they are touching or somehow entangled with other objects.
Building and programming a robot that is able to recognize the sunglasses in any possible configuration and then pick them up, fold them and put them back in their case is so difficult that we can probably conclude that the housekeeper’s job is relatively safe for the time being.
Cloud robotics is likely to be a powerful tool in ultimately solving that challenge. Android phones already have a feature called “Google Goggles” that allows users to take photos of an object and then have the system identify it. As this feature gets better and faster, it’s easy to see how it could have a dramatic impact on advances in robotics. A robot in your home or in a commercial setting could take advantage of a database comprising the visual information entered by tens of millions of mobile device users all over the world. That will go a long way toward ultimately making object recognition and manipulation practical and affordable.
In general, there are some important advantages to the cloud-based approach:
- As in the object recognition example, robots will be able to take advantage of of a wide range of online data resources.
- Migrating more intelligence into the cloud will make robots more affordable, and it will be possible to upgrade their capability remotely — without any need for expensive hardware modifications. Repair and maintenance might also be significantly easier and largely dealt with remotely.
- As noted in the video below, it will be possible to train one robot, and then have an unlimited number of other robots instantly acquire that knowledge via the cloud. As I wrote previously, I think that machine learning is likely to be be highly disruptive to the job market at some point in the future in part because of this ability to rapidly scale what machines learn across entire organizations — potentially threatening huge numbers of jobs.
The last point cannot be emphasized enough. I think that many economists and others who dismiss the potential for robots and automation to dramatically impact the job market have not fully assimilated the implications of machine learning. Human workers need to be trained individually, and that is a very expensive, time-consuming and error-prone process. Machines are different: train just one and all the others acquire the knowledge. And as each machine improves, all the others benefit immediately.
Imagine that a company like FedEx or UPS could train ONE worker and then have its entire workforce instantly acquire those skills with perfect proficiency and consistency. That is the promise of machine learning when “workers” are no longer human. And, of course, machine learning will not be limited to just robots performing manipulative tasks — software applications employed in knowledge-based tasks are also going to get much smarter.
The bottom line is that nearly any type of work that is on some level routine in nature — regardless of the skill level or educational requirements — is likely to someday be impacted by these technologies. The only real question is how soon it will happen.
The video below is a presentation from Google’s I/O conference on Cloud Robotics. It is fairly long and very technical, but if you have a strong interest or would like to see what some actual robot programming code looks like, check it out: