[Editor’s Note: This post was revised on 1/28/2016 on Ben’s request. See his note below.]
Author’s Note: Since its initial publication, I have reframed this post to more fully integrate the argument and data. This revised post reflects these changes.
Recent years have brought a resurgence of interest in how the rapid evolution of computer technologies is affecting work. Some have examined how smart machines are replacing manual labor, swallowing up the manufacturing jobs that have driven the growth of China’s economy. Others reveal how algorithms are supplanting knowledge workers. “Big data” and “machine learning” techniques help software engineers create algorithms that make more accurate and less biased judgments than well-trained humans. Software is already doing the work of medical lab technicians and replicating higher-order cognitive functioning, such as detecting human emotions and facial expressions, processing language, and even writing news articles.
Technology has long played a role in both eliminating certain types of work and creating new opportunities. Today’s debates often echo those of the past: technophiles believe that “disruption” is a source of social progress, whereas detractors worry that the coming waves of automation will deepen the insecurity and exploitation of workers. Both sides, however, often overlook the surprising ways in which, rather than creating “frictionless” economies, automation can in fact intensify the use of human labor.
In the remainder of this piece, I compare an exemplary study of the industrial revolution of the 19th century with a case study from the front lines of the automation revolution that many believe is now underway. In the Victorian era, new machinery did not replace human workers, but in fact often expanded their use. The same was true at a tech startup that I observed, where artificial intelligence was combined with the routinized application of human labor. Both of these cases draw attention to the specific ways in which technology restructures labor markets not only by eliminating jobs, but also by creating new types of work that must keep pace with machinery.
A View from the Workplace
In 2012 and 2013, I conducted participant-observation research at a startup company that I’ll call AllDone, which administered an online market for local services in the United States. AllDone’s strategy was to use advances in computer technologies to connect buyers and sellers of local services—from wedding photographers and DJs to electricians and plumbers—more efficiently than ever before. The process appeared simple: consumers would visit the website and enter a few details about the job that they wanted done; the company would then introduce them to local providers of that service who might be interested in bidding for the job. Sellers paid AllDone a small fee for each completed introduction.
If this description manages to suggest a symphony of elegant algorithms seamlessly connecting buyers and sellers, the organization of work that I witnessed at AllDone bore a striking resemblance to a 19th-century British factory. In 1977, historian Raphael Samuel published a masterful account of how the industrial revolution affected jobs in mid-Victorian Britain. Whereas scholars had traditionally believed that the technological innovations of the 19th century supplanted handicraft skill, Samuel presents a plethora of firsthand accounts of manufacturing work to demonstrate that, rather than rendering human labor unnecessary, the industrial revolution “created a whole new world of labour-intensive jobs” (8). In industries as diverse as mining, agriculture and food production, the building and construction trades, woodworking, and metallurgy, a familiar pattern emerged: as soon as a new piece of equipment was introduced into factories, new jobs would spring up in and around the machine.
Just one of Samuel’s countless examples is drawn from the mining industry. He describes how new steam-driven fans and pumps allowed workers and engineers to reach deeper deposits of coal. But larger mines meant that additional haulers were needed to handle more coal, while “longer galleries to travel…meant more roofs to prop, more roads to keep up, more rails to be laid down, while the increased use of blasting meant more hand-bored holes” (21). Rather than replacing workers, machinery could even give rise to entirely new occupational positions. In fact, manufacturers often preferred workers to machines because their capabilities and cost made them more attractive than new equipment. In addition to delivering more reliably high-quality output, workers tended to be cheaper, more adaptable, and more easily replaced. Technological change had indeed ushered in an era of unprecedented economic expansion. However, this growth was driven in part by how the application of innovative technologies expanded the division of labor and created a superabundance of low-cost workers.
I observed something similar at AllDone. About 20 employees worked in AllDone’s San Francisco office, half of them the software engineers and designers who built and managed the company’s technological infrastructure. But the San Francisco staff was dwarfed by the company’s online, work-from-home team of 200 contractors spread throughout the Philippines.
Just like in Samuel’s factories, at AllDone, “machinery was rarely self-acting, but required skilled hands to guide and to complete its work” (52). AllDone’s computer algorithms were supported by what some scholars and practitioners call “artificial artificial intelligence” (AAI) or “human computation,” whereby software engineers embed routine human work into computational infrastructure. Rather than tasking its small and expensive engineering staff with perfecting the website’s technological infrastructure, AllDone found tremendous value in recruiting humans to perform thousands of discrete, routinized tasks every day in and around its algorithmic machinery.
AllDone relied on human labor to supplement its code base for the same reasons that 19th-century factory owners often preferred workers to equipment: flexibility and low cost. What were some of the tasks that this human computation was used for? Like many other companies, AllDone uses search engine optimization (SEO) techniques to try to bump its pages to the top of search engine results. Companies that want to get in front of consumers consequently create pages that are rich in the “keywords” that consumers search for (e.g. “best locksmith,” “affordable tutor”). AllDone’s engineers used 50 members of AllDone’s team in the Philippines to help the company attract buyers. They set up a web portal that would show writers descriptions of its sellers and the top keywords for the services they offered. Every month, team members wrote around 50,000 pieces of keyword-rich content (i.e. descriptions of the services that sellers offered) that could then be added to AllDone’s pages.
The low cost of hiring workers in the Philippines, where the prevailing wage for qualified workers is far lower than in San Francisco, allowed the company to undertake projects that otherwise would have been infeasible. AllDone’s engineers set up a process that allowed team members in the Philippines to curate each new seller’s profile to enhance the perceived trustworthiness of the company’s offerings. For instance, members of one team eliminated sellers whose services violated the company’s guidelines, verified professional licenses in online state databases, ran voluntary criminal background checks, and even proofread the text that sellers entered in the profiles to ensure a minimum standard of professionalism. They also boosted sellers’ trust in AllDone’s buyers by manually vetting each consumer request before distributing them to sellers, throwing out any that looked fraudulent (e.g. “Mickey Mouse of Orlando, FL wants to hire a dog trainer for Pluto”).
AllDone’s remote teams were also more flexible than computer algorithms. AllDone maintained two dozen contractors on a “special projects” team who were available to handle spur-of-the-moment missions. Imagine that a company wants to undertake a data mining project to gather a massive amount of information about potential users or competitors. Some would assign engineers to code up smart programs to scrape that data from the web. AllDone could instead “throw bodies at the problem,” deploying human computation to undertake the project on a moment’s notice, and increasing the pace of innovation.
Finally, AllDone’s remote teams could be “scaled up” or “scaled down” to meet fluctuating user demand. Rather than perfecting an algorithm to match consumer requests with the appropriate sellers, AllDone maintained a team in the Philippines to manually construct each connection. As request volume grew, the team simply continued to expand, until nearly 100 workers had been trained on the process and were available to link each of thousands of buyers with dozens of sellers per day, 24 hours a day, seven days a week. Ceding the process to human computation freed up engineers to test new, revenue-generating projects that would be crucial to the company’s growth. This is why, like Samuel’s factory owners, AllDone continued to “rely on workers’ skills even when there was machinery ready, in principle, to replace them” (47-8).
What Kind of Future?
Steam power and hand technology “may well appear as belonging to different epochs, the one innovatory, the other ‘traditional’ and unchanging in its ways,” Samuel wrote, but instead “they were two sides of the same coin” in 19th-century Britain (57). The same could be said of the information technology and hand work used by AllDone in the early 21st century.
Since I left AllDone, the company’s financial resources and engineering staff have expanded dramatically, along with its user base. Meanwhile, the team in the Philippines has more than doubled in size. While some processes have been automated, new tasks have been added to its purview. AllDone’s top managers recognize what Samuel’s factory owners saw 150 years ago: that high-tech tools can be more powerful when they’re interwoven with high-quality, low-cost human labor that is flexible and scalable.
Though the similarities are striking, we can also observe important differences in how technology restructures labor markets and affects the subjective experience of work across time. By de-skilling work, British factories in effect created the same low-wage workers upon whom they would subsequently rely to complement their new technologies. Those who had previously been craft workers experienced this process as the degradation of their labor. The situation today is different: as software developers in rich nations automate certain tasks, they open up new sources of, and demand for, low-cost labor around the globe. Consequently, much of the work in and around today’s new machines can be performed remotely by workers in developing countries whose access to alternative sources of income may be limited. AllDone’s Filipino team members frequently reported that they preferred working from home for a startup to commuting to high-pressure jobs in outsourced call centers. For many, the subjective experience of high-tech hand work is thus intertwined with the promise of new opportunities in a global economy.
Work is not only changing for the people whose efforts complement algorithms, but also for those tasked with creating the software itself. Instead of being limited to experimenting with code, software engineers are increasingly “tinkering” with human labor. As Shreeharsh Kelkar notes, today’s computer scientists are encouraged to “concentrate on designing useful assemblages of humans and computers rather than on creating intelligent programs.” Tech giants like Google, Facebook, and Twitter use online contractors to rate the search engine results produced by algorithms, filter out inappropriate user-generated content or advertising, and sort content to identify trending topics. Netflix has combined the power of software with the cognitive capacities of human viewers to create an innovative categorization scheme that classifies content into nearly 77,000 micro-genres. In these and many other cases, innovations in programming are not premised on generalized AI that delegates “thought” to machines. Instead, code is developed to accomplish specific tasks, with the knowledge that its execution will require human input, interpretation, or supervision.
Both 19th-century capitalists and AllDone’s managers operated in environments in which consumer demand, products, and production techniques were in flux, and both had access to relatively cheap and flexible sources of labor. Given these circumstances, both found it advantageous to supplement innovative technologies with a great deal of “traditional” routinized human labor. It is likely that AI will continue to fall short of replicating human cognition for the foreseeable future. Many tasks—including vetting content and gathering and standardizing data from disparate sources—require cultural competencies that remain expensive and time-intensive to program. Meanwhile, the supply of AAI will increase as global, online labor markets grow.
In the coming years, many of the innovations emerging from high-tech hubs like Silicon Valley will be enabled by the efforts of far-flung, flexible, and relatively low-cost workers. Those who focus on how automation eliminates jobs often miss how it can also give rise to new types of work that may be hidden from view. By examining the objective conditions and subjective experiences of work in emerging global software production networks, analysts can help societies understand and respond to these new realities.
Samuel, Raphael. 1977. “Workshop of the World: Steam Power and Hand Technology in mid-Victorian Britain.” History Workshop Journal 3(1): 6-72.
Benjamin Shestakofsky is a Ph.D. candidate in the Department of Sociology at UC Berkeley. He is interested in the sociology of work, economic sociology, online markets, and organizational culture.