Are smart machines coming for our jobs?
In the past, technological change has generally led to the displacement of workers from some jobs, but also to the creation of new work. For example, as automation reduced the number of workers needed to grow and harvest crops in the early 20th century, technological change resulted in employment gains in the manufacturing and service sectors.
Today, however, many worry that the historical link between technological innovation and job creation may be coming to an end.
Big data and artificial intelligence make it possible for computers to perform tasks that previously required complex human cognition. Software algorithms are already driving cars, diagnosing diseases, and writing news articles.
A credible case can be made that, thanks to the rapid development of AI, this wave of technological change will usher in an era of widespread unemployment.
Most contemporary inquiries into the future of work offer projections of employment trends at the level of industries or occupations. These studies are useful for helping us conceptualize broad shifts in labor markets, but they aren’t able to shed light on the complex and unpredictable ways in which human workers and software systems interact in real-world settings.
In a recent study, I argue that in-depth examinations of the organizations in which software algorithms are developed and implemented can help us generate new insights into the question of when software systems function autonomously, and when they rely on the assistance of complementary human workers.
To investigate this question, I spent 19 months inside a San Francisco-based tech startup operating at the frontiers of the digital economy. I discovered how software automation can generate new and surprising forms of human intervention and employment.
When computers alone were unable to complete an operation, the company mobilized workers located across the Philippines to provide computational labor, completing routine information-processing tasks to support or stand in for software algorithms. Workers in the Las Vegas area performed emotional labor aimed at helping users adapt to changing software systems.
Instead of perfecting artificial intelligence that would progressively push people out of the production process, managers continually reconfigured ensembles of technology and human helpers.
The Amazon of local services
I conducted participant-observation research at a startup company I call AllDone, which aimed to become “the Amazon of local services” by making it as easy to hire someone to do a job for you as it is to buy a book or a toothbrush online. AllDone’s marketplace covers hundreds of service categories ranging from house cleaners and math tutors, to wedding photographers and DJs, and far beyond.
Buyers who visit AllDone’s website place a request by specifying the details of the job they want to hire someone to do for them. Then the company sends the request to its network of local sellers, who pay AllDone a small fee to send the buyer a message pitching their services.
Working alongside members of the organization allowed me to observe three distinct periods of the company’s development, each lasting roughly six months. During each phase, the company recalibrated its strategic direction to meet the expectations of the investors who funded the firm’s growth.
Each strategy generated new organizational problems for managers to solve. Managers addressed these problems by creating or expanding configurations of software systems and human workers in global production networks. Below I provide a sketch of human-machine interactions during each of the three phases I observed.
Phase I: Expansion
Following its first round of funding from a venture capital firm, AllDone’s leaders planned to use the cash infusion to hire more software engineers and accelerate product development.
Engineers spent the majority of their workdays reviewing job applications and interviewing candidates, leaving them little time to manage and update the systems that made the website work.
When we hear the word “algorithm,” we usually think of a computer transforming information according to rules specified by a programmer. At AllDone, however, human workers were often mobilized to implement information-processing tasks by hand to support or stand in for software systems.
To keep AllDone’s website running while developers were preoccupied with recruiting activities, the company expanded its digital assembly line in the Philippines, where workers performed computational labor to complement technology.
Instead of building software algorithms to connect buyers with sellers, AllDone’s team of developers passed the matching process along to Filipino workers who logged into an administrative portal through which they manually constructed every introduction.
Users never knew that a person—rather than a computer algorithm—had handcrafted each introduction. Workers in the Philippines facilitated the company’s growth by completing hundreds of thousands of computational tasks each month.
Phase II: Revenue Generation
After proving to investors that lots of people wanted to use their product, AllDone now had to demonstrate that they could actually make money off it.
The pivot toward boosting revenue presented AllDone’s leaders with a new problem. AllDone’s sellers—the service providers who paid the company to send their quotes to buyers—constituted the company’s sole source of income. AllDone would have to retain the business of its most active sellers at the same time that it dramatically increased the fees they paid to use the product.
AllDone addressed this problem by again combining its technological infrastructure with complementary human workers. The company built a phone support team in the Las Vegas area to perform emotional labor aimed at helping users adapt to the new system. These frontline workers regulated their emotional display in an effort to instill feelings of trust in AllDone’s customers.
AllDone’s phone support team was tasked with calling AllDone’s most active sellers to let them know about the impending changes to the fee structure. Support agents would often absorb a barrage of outrage and insults from sellers who typically accused the company of being greedy, or ruining their livelihoods, or worse.
The agents used their intuition, creativity, problem solving skills, and powers of persuasion to solve problems of trust in interpersonal interactions and convince sellers to continue to use the platform in spite of the higher fees. These functions were nearly impossible to automate, yet integral to the operation of AllDone’s software and business.
Phase III: Expansion + Revenue Generation
With its new business model in place and its next round of funding secured, AllDone combined its prior goals of expansion and revenue generation. Its developers began to implement machine-learning algorithms to automate a variety of tasks that had previously been performed by workers, including the matching process described above.
But in spite of this new wave of automation, the company continued to rely on human workers to complement its software. Even as AllDone’s engineering corps expanded exponentially, the size of its remote teams more than quadrupled.
As the firm matured, developers continued to confront the same problems that they had previously faced. Software alone was still incapable of keeping up with developers’ needs, and users continued to struggle to adapt to novel systems.
AllDone’s architects were repeatedly faced with technology’s limitations. As the ranks of software engineers grew, their increased pace of innovation necessitated the continual intervention of complementary workers to bring technology up to speed with developers’ vision, and to help users adapt to changing systems.
An Organizational Perspective on Technological Change
Existing studies of the relationship between software automation and work largely center on macro-level phenomena like labor markets, job categories, and general work tasks. My research reveals processes through which, at the level of the firm, complementary labor is continually incorporated into and displaced from software systems.
In an age of so-called disruption, the internal dynamism of organizations like AllDone is likely to make human labor an integral part of an increasingly digitized economy.
The outcomes of technological innovation are thus unlikely to be monolithic, but will instead depend on the contexts in which software development is applied and the rate at which those contexts change.
My study suggests that the rate of organizational change will be an important determinant of the shape of human-software configurations. In firms in which software products remain relatively stable, we’re more likely to see the substitution of software for workers outpace the creation of complementary jobs.
The future of work is inextricably linked to the future of organizations. Close examinations of the social settings in which software algorithms are developed and implemented will yield important insights into the consequences of rapid technological change.