Predicting Human Obsolescence, One Job at a Time

Well, a couple of Oxford professors may have an answer for you. In an influential 2013 paper called "The Future of Employment: How susceptible are jobs to computerisation?", authors Carl Benedikt Frey and Michael A. Osborne concluded that 47 percent of U.S. jobs were at high risk for computerized substitution within "some unspecified number of years, perhaps a decade or two." Not only that, they devised a formula for analyzing 702 specified jobs and assigning each one an individual computerization score between zero and one. The higher the score, the greater the risk of automation in the near future. For example:

Frey and Osborne observe that in the past, machine substitution of human labor has taken place almost exclusively in occupations with "routine tasks involving explicit rule-based activities." Another way of putting this is to ask yourself the following question: Can the job be easily described in a clear list of instructions that are repeated? Think of many telemarketing operations:

Another example would be repetitive assembly line labor, where the worker welds together the same two pieces on an unending procession of identical automobile doors. These kinds of jobs are what economists might call "routine intensive occupations," and if there's a job like this that hasn't already been taken over by a machine, it is in imminent danger of automation in the near future.

However, while only the most routine tasks became machine fodder in previous decades, Frey and Osborne point out that recent developments in big data, machine learning and mobile robotics mean that machines are now able to perform both cognitive and manual jobs that people once thought were relatively immune from the machine invasion. 

To illustrate this, Frey and Osborne quote a 2003 paper from The Quarterly Journal of Economics, in which the authors (Autor, Levy and Murnane) write, "Navigating a car through city traffic or deciphering the scrawled handwriting on a personal check – minor undertakings for most adults – are not routine tasks by our definition." Today, Google's autonomous fleet has demonstrated pretty soundly that cars without human drivers are safer than cars with them, and depositing a handwritten check by taking a picture with your smartphone is a commonplace. These are specific examples of a general trend: Jobs that used to seem like they couldn't be performed by a programmatic software routine not only can be, but in many cases already are.

Frey and Osborne's paper is quite interesting and worth a read if you'd like to learn more about the methodology they use to come up with these risk assessments, but the simplified version is that lower computerization scores went to jobs that require key skills that remain the most difficult for computer-based machines. These difficult-to-automate skill categories include:

It's worth noting that Frey and Osborne don't express the belief that these skills are in principle inaccessible to machines. Instead, they claim that they will take longer to attain because of "engineering bottlenecks," meaning we simply don't yet have the knowledge or technology that would allow us to program them, so these skills probably won't be substituted by computer capital in the next decade or two.

After analyzing for these criteria, it appears that the safest areas of employment are management, education, health care, arts and media, engineering and science. Specific examples of jobs that rate very low on their computerization score are:

The employment sectors most at risk are transportation and logistics, office and administrative support workers, manufacturing and production, and service occupations. Some examples of jobs that rated very high on the computerization score are:

Keep in mind, however, that as well-informed as these assessments may be, Frey and Osborne point out that humans are not always very good at predicting the extent to which something can be automated (remember those autonomous cars and check-cashing examples?).

Plus, there could be other, unknown factors applying pressure in the opposite direction, causing us to overestimate machine capabilities and underestimate the value provided by human workers. For example, think about human versatility, or the ability to do a wide range of different and often unexpected tasks well.

Every good robot in the world is a specialist robot. It is good at doing one human job, or, at most, a handful of well-defined jobs. There is no such thing as a good generalist robot, able to do every physical and mental job a human does with reasonable success. Not only is there no such robot, we're not even close.

Watching a well-trained industrial robot repeatedly executing its singular job can be hypnotic and intimidating. The welding arms lining the assembly chain of an auto manufacturing plant move with startling grace and speed. But that's exactly because they have one job, and one job only. Watching robots try to fulfill diverse physical specifications is another matter entirely. For example, take a look at the delightful robots designed to complete the 2015 finals for the DARPA Robotics Challenge, which specifically encourages physical versatility in robot design by requiring the robots to perform several variegated types of locomotion and physical manipulation, like walking up some stairs, opening a door, turning a valve and navigating rubble.

Those robots you see crumpling into heaps when defeated by a doorknob or some sandy terrain are designed by extremely smart people who know what they're doing. The repeated failures of the robots in the competition are not indicative of bad roboticists and engineers, but of the huge difficulty cramming lots of physical versatility into a single machine.

With this in mind, let's think about the average shift of a human restaurant server. You have to take orders, answer questions about the menu, recognize food orders and which tables they correspond to, transport food from the kitchen to the tables, transport dirty dishes from the tables to the dishwasher, clean up spills and dropped items, answer special requests ("Can you make this pizza without dough?", "My child threw his fork on the floor and needs a new one ..."). And then there are a thousand other small tasks one wouldn't even think about, like lighting candles on a table, recognizing and substituting improperly cleaned dishes and cutlery, or recognizing regulars and chatting with them.

Now that doesn't mean it's impossible to automate a restaurant. In fact, some have done it.

We'll let you judge whether you'd find that experience appealing, or not so much. Either way, it would require lots of planning, huge capital investment and a fundamental alteration of the restaurant experience. And that last concern might be key: What if people don't want to go to a restaurant with 14 specialized robots instead of a human server? What if this fundamentally reduces the value of the experience the restaurant is selling?

In addition, developing these kinds of robots is difficult and expensive, and economic pressures will win out. If you look at the example of food service in the United States, many restaurant servers are paid extremely low cash wages (often something like $2.13 an hour) under the rationalization that they will make up the difference in tips, voluntarily given by the customers. At such a low price, there's not an incredibly strong incentive for restaurants to replace servers with robots, especially the clumsy, dish-smashing early prototypes that will hit the market first.

The question of impending automation in cases like this is ultimately an economic one. It's not always a case of whether it's merely possible to create a robot to do a job, but whether the value minus cost that the robot provides exceeds the value minus cost of the human laborer. It may be in many such cases that the hidden value of human labor lies.