home


~Automation and the Non-workforce~
the future of labor in the robotics age

10/2015, J. Haupt


3/2017 update: I now consider this conclusion flawed. See note on home page.



The Industrial Revolution and the Information Age sponsored economies that continued to create jobs in response to technological unemployment, but as we approach a robotics age it may be fool-hardy to assume this will continue. In a world where machines have become sufficiently advanced to replace human workers in many capacities, what would be the nature of the remaining work and what would become of those unsuited to it? The answer may have more bearing on the long-term well being of humanity than any of the enduring questions of our time. Confronting this degree of technological unemployment, the world’s economies would need to adapt to support an ever-increasing sector of people with little hope of having a productive livelihood. More importantly, a cultural shift of the place of unemployment would be needed. In the face of unavoidable job loss and even foreseeing the potential beauty in a society where the burden of needing to work has been rendered unnecessary, unemployment should eventually be considered a sign of progress, not retrogression.

This page makes a prediction and assesses the effect of that prediction. The prediction is that technology, mostly in the form of robotics, will have the capacity to replace human workers in most current roles in the foreseeable future. An overview of current state of the art robotics research and a review of the rate of commercialization of automation technologies may convince the reader that this is likely, or at least a valid possibility. A plausible effect of this prediction is that job loss, rather than job displacement, will prevail in the 10-30 year timeframe. The initial mechanization of society and the subsequent degrees of automation that came with the computer age have undoubtedly created jobs, but in the case of a technology designed with the specific intent to reduce the need for human effort, it seems likely that job creation will be unable to keep up with automation-resultant job loss and force an ever larger portion of the population into permanent unemployment.


Artificial intelligence is improving

Artificial intelligence (AI) has improved to the point of being able to do things that only human intelligence could do just a few years ago. Machines will soon be able to operate vehicles of every kind (cars, aircraft, trains) without human intervention (Gannes, 2014; Markoff, 2015;  Jaffe, 2015), and as described below enabling technologies are maturing that will allow machines to replace many other jobs. Bankers, store clerks, cleaning service workers, civil servants, construction workers, machinists, custodians, law enforcement officials, anyone operating a machine or vehicle, assembly workers, and even tradesmen and various professionals are at risk in the coming decades. 

Until recently,  a division existed between those tasks that machines are classically good at, like rules-based problems, repetitive automation, etc. and those tasks that machines are classically bad at but which humans find easy, like object classification, spatial reasoning, planning, language, etc. This barrier was an accepted quality of computer capability that didn’t seem likely to change. Levy and Murnane (2004)  elucidate the distinction between these categories of problems in a 2004 book which concludes that the fear of job destruction by automation is misplaced because machines can’t handle human-like information processing, in particular “complex pattern recognition” (p. 30). These authors give examples of problems that machines of the day were not only incapable of handling but were thought unlikely ever to handle, like the extreme case of a driver navigating a busy road: “... executing a left turn across oncoming traffic involves so many factors that it is hard to imagine discovering the set of rules that can replicate the drivers behavior.” (p.20) Only 10 years later, this problem has been solved, with limited self-driving functionality already available on current models and the first fully autonomous models likely to be released in the next 5 years (Greenough, 2015). Heavy equipment and mining companies are also on the automation bandwagon. Caterpiller has a self-driving mining truck well into development and the company appears to be betting on autonomous equipment in general (Ortt, 2013).

The case of self-driving cars should not be seen as isolated, with application only to the particular case of a car on a road. The underlying power of this technology speaks to a broader world and the computational methodology can be applied to most human-dominated tasks. Indeed, applying this new-found artificial intelligence might to even more impressive skill sets – those requiring fine dexterity --  is well under way. With particular regard to domestic robotics, research groups have made strides designing robots that do things like folding towels (Maitin-Shepard, Cusumano-Towner, Lei, & Abbeel, 2010), tying knots (Schulman, Ho, Lee, & Abbeel, 2013) and doing laundry (Srivastava, Zilberstein, Gupta, Abbeel, & Russell, 2015) autonomously. Eric Brynjolfsson and Andrew McAfee (2014) encapsulate the impact of these developments nicely:

Progress on some of the oldest and toughest challenges associated with computers, robots, and other digital gear was gradual for a long time. Then in the past few years it became sudden; digital gear started racing ahead, accomplishing tasks it has always been lousy at and displaying skills it was not supposed to acquire anytime soon (p.20).

The idea of a self-directed robot that can dig ditches, pave roads, or build a house should not seem far-fetched, and robots specializing in human interaction are already arriving. Pepper, a humanoid social robot by Aldebaran Robotics, and Q.bo, an open source robotics platform by The Corpora S.L. are examples of this type that are available now. Pepper and Q.bo are both technology concept products, with the Q.bo in particular marketed to developers and early adopters. These are relatively un-groundbreaking machines having leveraged the sci-fi-esque voice recognition and language processing technology already available on smart phones, but they represent an important and necessary integration of human interaction with mobile robotics platforms and will be an important component of the more functional robots to come. This flurry of AI development coupled with the increasing integration of smart technology into day-to-day life leads many to conclude that we’re on a natural course for a robotics age (Wadhwa, 2015; Kalin, 2015).

What changed in the past 10 years that made human-like interpretation of the world go from being all but impossible to being commercialized? Several groundbreaking development happened in the 2009-2013 timeframe to a particular application of artificial intelligence programming called neural networks. Thanks in large part to a computer scientist named Geoffrey Hinton, these efforts culminated in a revolutionary framework for artificial intelligence implementation known as Deep Learning (Markoff, 2012).  To respond to the quote from Levy and Murnane above about it being “...hard to imagine discovering the set of rules that can replicate the drivers behavior”, with the advent of Deep Learning the game has changed so that a “set of rules” that can account for every possible eventuality in a given scenario, like driving a car, is no longer needed. Deep Learning has supplanted traditional rules-based design in AI systems so that problems are processed in an organic way, modeled after the natural animal brain to process information as a mind does. The result is a system which isn’t programmed in the conventional sense, but trained. 

Only computer scientists could have predicted what the general public now takes for granted: smart devices with the intelligence to recognize friends in photos and answer verbal questions (capabilities also being adopted by companies eager to reduce call center costs). This power extends to the ability of computers to write grammatically flawless, well conceived prose. The Associated Press has begun exercising this ability by assigning the task of writing certain news blurbs to software (Miller, 2015). The breathtaking commonality and accessibility of these technologies has a way of making them seem mundane, but there should be no misconceptions about what has happened here. The current state of object recognition and language processing, previously “holy grails” of computer science, is nothing short of miraculous. Even still, this is only the earliest fruit of Deep Learning.

Perhaps the first sign that the robotics age is beginning is an industrial robot called Baxter. Rethink Robotics’ Baxter is a $25,000 anthropomorphized automaton marketed for assembly line work, and true to form in the era of the Deep Learning algorithm it is trained to perform its tasks, not programmed (Kelly, 2012). An employer needing different, oddly shaped parts placed side by side into a box needs only to guide Baxter’s hands through the motions of the task, and “he” takes it from there. The robots skill isn’t merely in replicating the course arm and hand motions it recorded during the training phase, but (in the example of a pick and place task), interpreting new objects in different places, recognizing the pattern with which they need to be arranged, and having the dexterity to arrange the objects even though training may have involved only one or two iterations (Knight, 2012). Baxter is both an amazing testament to modern technology and a humble prelude to what soon will be. As the cellular telephone market changed society by the introduction of viable smart phones, the enabling technologies being developed by academia will surely mature into something even more amazing than Baxter.

A good case study of a company making a proactive effort to automate its workforce is Amazon. So far, 10 of Amazon’s fulfillment centers throughout the United States have been upgraded with workforces of Kiva fulfillment robots (Tam, 2014). These warehouses are veritable dance floors for a dizzying metropolis of robotically moved merchandise, and they’re a pioneering example of a highly automated distribution environment. Many businesses have similar distribution needs, like wholesale clubs, super markets and lumber yards, and with the technological hurdles collapsing one wonders what the adoption timeline will be in these other sectors. It’s no secret the company is also putting development effort into a drone-based package delivery system (Halpin, 2015). With their current automation success in internal fulfillment one wonders if Amazon can start treading on the territory of parcel delivery services.

Clearly, the disruptive power of the information will not be confined to the virtual world. It’s becoming apparent that jobs like taxi and bus drivers, truck drivers, pilots, railroad engineers, factory workers, warehouse workers, call center operators, and even low-level journalists are poor options for today’s newborns, and it seems other realms of work will be at similar risk as bleeding-edge AI and robotics research grows to commercialization.


Technological unemployment

Which are the kinds of work least at risk from technology? Work involving intrinsically human pursuits, like philosophy and all forms of art, are difficult to imagine being replaced. Continuing with the least at-risk jobs, scientific pursuit, engineering, literature, music, advertising, law, management, executive business, and politics also seem safe, barring more fundamental changes. Can the entire workforce be encompassed only by these employment categories? What most of those jobs have in common is that they demand high education, which immediately excludes the portion of the population that’s unwilling or unable to attain to intellectually intensive livelihoods.

On the other side of the spectrum we have the jobs at high risk: Those for which relevant automation technologies are already maturing like vehicle operators and factory workers. Will machines make viable replacements for wait staff in the next 30 years? Or line cooks? Possibly, possibly not. What about landscapers, construction workers, electricians, plumbers, and cleaners? Considering robots can now autonomously tie knots and hold conversations but only 10 years ago couldn’t navigate a room or parse the most basic sentences, automating this kind of work is plausible.

According to the U.S. Department of Labor (2014) the sectors currently employing the most people are state and local government, professional and business service, and healthcare and social assistance; some of the jobs least at risk for technological unemployment which together account for 37.1% of the workforce. Also excluding educational services and the federal government as low risk sectors, if a modest incarnation of the predictions of this essay play out and the admittedly wild guess that half of the remaining sectors are at high-risk for automation, the labor force could expect a sizeable permanent reduction. Undoubtedly there will be other employment sectors that absorb some of these jobs, and undoubtedly new industries will appear, but the likelihood that all or most of the labor force can be absorbed in this way seems low.

Predictions aside, technological unemployment may already have begun. Brynjolfsson and McAfee quantify what they call bounty (the availability and range of inventions and services), and spread (productivity, employment, and income distribution), using these concepts to iterate that technological unemployment is already well under way. Two metrics in particular, productivity (company earnings relative to salaries paid) and employment, which tracked each other for many years have become decoupled in the past 20. They write:
While these two key economic statistics tracked each other for most of the postwar period, they became decoupled in the late 1990s. Productivity continued its upward path as employment sagged. Today, the employment to population ratio is lower than any time in at least 20 years, and the real income of the median worker is lower today than in the 1990s.  Meanwhile, like productivity, GDP, corporate investment, and after-tax profits are also at record highs (p. 164).


Adapting

Various modes of adaptation are possible as unemployment increases and becomes recognized as non-reversible. A general, non-exhaustive summary of some such modes are enumerated below:

1.    Economic mechanisms encourage the formation of non-essential jobs.
2.    Policy is drafted to limit the adoption of automation technology.
3.    New and unexpected jobs are created, allowing employment to recover.
4.    Social programs expand.

Scenario 1 above is an unlikely total solution to technological unemployment because companies are naturally driven to improve the ratio of productivity to cost, but some degree of artificial job formation may play a role in government sectors where the access of politicians motivated to improve employment statistics can result in increased public works funding and inflated bureaucracies.

Scenario 2 would be the most woesome reaction. I again reference Brynjolfsson and McAfee for an excellent summary of the problem here:

Doing so would be about as bad an idea as locking all the schools and burning all the scientific journals. At best, such moves would ensure the status quo at the expense of betterment or progress. As the technologist Tim O’Reilly puts it, they’d be efforts to protect the past against the future. So would attempts to protect today’s jobs by short-circuiting tomorrow’s technologies. We need to let the technologies of the second machine age do their work and find ways to deal with the challenges they will bring with them (p. 231).

Scenario 3 is the non-scenario. It is entirely possible that technology will not lead to permanent job loss, and that current economic trends that would seem to be indicators of the beginning of technological unemployment are being cause by other factors, like globalization. A common thread in economic thinking is that there’s substantial historical grounding for the belief that technology will always create jobs, or more severely, that true technological unemployment is downright impossible. Brynjolfsson and McAfee offer a thought experiment to prove that there isn’t an “iron law” that forbids technological unemployment by imagining a technology that could cause it: They imagine the invention of an android “...that could do absolutely everything a human worker could do, including building more androids.” (p. 180) They go on to describe the profound economic effects; the massive increases in productivity, decreases in prices, and widespread unemployment. “Every economically rational employer would prefer androids, since compared to the status quo they provide equal capability at lower cost... Entrepreneurs would continue to develop novel products, create new markets, and found companies, but they’d staff these companies with androids instead of people.” (ibid) A natural response to this is that the android described is as hypothetical as the argument itself. But the purpose of a thought experiment is to consider an extreme circumstance to answer whether nature forbids a given scenario outright, we can at least see that the idea of technological unemployment as being fundamentally impossible is invalid. Of course, if the technology could be created to fully replace human labor, human labor would become obsolete.  

In scenario 4, whether or not technological unemployment is identified as the cause, it follows that if unemployment benefits for the poorest classes expand to include a much larger portion of the population the economy could be seen as having adapted successfully. The cultural shift needed here is more fundamental than this. Society might continue to stigmatize unemployment, but if unemployment can be recognized publically as an inevitable outcome of progress then the stigma attached to being unemployed will give way to acceptance. Rightfully so, because acceptance of the reality of permanent unemployment would beget a cultural satisfaction in having won a society that simply doesn’t require a high (or even a moderate) employment rate in order to thrive. To quote Arthur C. Clarke, “The goal of the future is full unemployment, so we can play.”

Brynjolfsson and McAfee offer recommendations for realizing scenario 4. One possibility is the so-called basic income, which would guarantee a fixed annual income to every citizen. However, they prefer a concept first proposed by the famous economist Milton Friedman called the negative income tax, which “combines a guaranteed minimum income with an incentive to work” (p. 238). All such proposals are based on a hybrid economy model that supports the unemployed via social programs while encouraging capitalism-driven entrepreneurism and competition on top of this socialist base. They underscore the importance of work to one’s sense of worth, and do not pretend that a truly labor-less economy is ideal. People want to work, and the opportunity to be able to work should always exist. It should also be pointed out that a reduced labor force doesn’t necessarily reveal itself in unemployment statistics. John Maynard Keynes was had the insight in 1930 that a workforce inundated with the bounty of technology might become satiated, a possible outcome being that most people would continue to work but might have 15 hour work weeks (as cited in Brynjolfsson and McAfee, p. 177). This seems like an ideal solution to the dualistic problem of people wanting to work (but not too excess), or of not being able to find enough work. Of course, an increase in the proportion of people working part-time alone would not bode well for the economy. We would need to have adapted so that people can work part-time and be happy doing so; a possible realization of the hybrid capitalist-socialist economy outlined by Brynjolfsson and McAfee. Whether the technological shift in employment is via increasingly part-time employment or via outright non-employment, the economic interventions listed above are possible roads to this end.

Artificial intelligence is now capable of doing things that only humans could do previously and emerging technologies are pointing to a future in which most jobs can be automated. The final steps to realizing this future are continued improvements to Deep Learning algorithms and a smart phone-like explosion of robotics products which can proxy conventional human labor. Economic indicators may already be telling us that technological job displacement is turning into technological job elimination. Preempting this scenario either by imposing restrictions on technology adoption or by creating superfluous jobs would be an artificial modification. Society will need to embrace the abilities of its technological progeny by moving from a mode in which unemployment is considered bad to one in which unemployment is the desired state of being.


References

Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of brilliant technologies. New York, NY: W. W. Norton & Company, Inc..

Floridi, L. (2014). Technological unemployment, leisure occupation, and the human project. Philosophy & technology, 27(2). http://dx.doi.org.ezproxy.snhu.edu/10.1007/s13347-014-0166-7

Gannes,  L. (2014). Here’s What It’s Like to Go for a Ride in Google’s Robot Car. Retrieved from http://recode.net/2014/05/13/googles-self-driving-car-a-smooth-test-ride-but-a-long-road-ahead/

Greenough,  J. (2015). THE SELF-DRIVING CAR REPORT: Forecasts, tech timelines, and the benefits and barriers that will impact adoption. Business Insider. Retrieved from http://www.businessinsider.com/report-10-million-self-driving-cars-will-be-on-the-road-by-2020-2015-5

Jaffe,  E. (2015).The case for driverless trains, by the numbers. Retrieved from http://www.citylab.com/tech/2015/04/the-case-for-driverless-trains-by-the-numbers/390408/

Kalin,  N. (2015).The robotics age is closer than we think. Huffington Post. Retrieved from http://www.huffingtonpost.com/natalie-kalin/four-items-that-should-be_b_6882640.html

Kelly,  K. (2012, December 24). Better than human: Why robots will – and must – take our jobs. Wired. Retrieved from http://www.wired.com/2012/12/ff-robots-will-take-our-jobs/

Knight, W. (2012, September 19). This robot could transform manufacturing. Technology Review. Retrieved from http://www.technologyreview.com/news/429248/this-robot-could-transform-manufacturing/

Levy, F., & Murnane, R. J. (2004). The new division of labor: How computers are creating the next job market. Princeton, NJ: Russell Sage Foundation.

Maitin-Shepard, J., Cusumano-Towner, M., Lei, J., & Abbeel, P. [RLLberkely]. (2010, March 17). (50X) Autonomously folding a pile of 5 previously-unseen towels [Video file].
Retrieved from https://www.youtube.com/watch?v=gy5g33S0Gzo

Markoff,  J. (2015). Planes without pilots. The New York Times. Retrieved from http://www.nytimes.com/2015/04/07/science/planes-without-pilots.html?_r=0

Markoff,  J. (2012). Scientists see promise in deep-learning programs. The New York Times. Retrieved from http://www.nytimes.com/2012/11/24/science/scientists-see-advances-in-deep-learning-a-part-of-artificial-intelligence.html?_r=0

Miller, R. (2015, January 29). AP’s ‘robot journalists’ are writing their own stories now. Retrieved from: http://www.theverge.com/2015/1/29/7939067/ap-journalism-automation-robots-financial-reporting

Minor, J. (2013, October 28). AI startup develops CAPTCHA-cracking software. In pcmag. Retrieved from http://www.pcmag.com/article2/0,2817,2426471,00.asp

Halpin,  N. (2015, June 22). Expect your Amazon deliveries within 30 minutes via drones next year. Business Insider. Retrieved from http://www.businessinsider.com/expect-your-amazon-deliveries-within-30-minutes-via-drones-next-year-2015-6

Rotman, D. (2013, June 12). How technology is destroying jobs. In MIT technology review. Retrieved from http://www.technologyreview.com/featuredstory/515926/how-technology-is-destroying-jobs/

Ortt, G. (2013, December 9). Caterpillar- Betting on autonomous machinery. Retrieved from: http://seekingalpha.com/article/1907581-caterpillar-betting-on-autonomous-machinery

Schulman, J., Ho, J., Lee, C., & Abbeel, P. (2013). Learning from demonstrations through the use of non-rigid registration. In Proceedings of the 16th International Symposium on Robotics Research (ISRR).

Srivastava, S., Zilberstein, S., Gupta, A., Abbeel, P., & Russell, S. (2015). Tractability of Planning with Loops. In Proceedings of the 29th AAAI Conference on Artificial Intelligence (AAAI-15)

Tam,  D. (2014). Meet Amazon’s busiest employee – the Kiva robot. CNET. Retrieved from http://www.cnet.com/news/meet-amazons-busiest-employee-the-kiva-robot/

United States Department of Labor Bureau of Labor Statistics. (2013). Employment by major industry sector [Data file]. Retrieved from www.bls.gov/emp/ep_table_201.htm

Wadhwa,  V. (2015). Welcome to the dawn of the age of robots. The Washington Post. Retrieved from https://www.washingtonpost.com/news/innovations/wp/2015/06/22/welcome-to-the-dawn-of-the-age-of-robots/