Hype, Profit, Labor, and Agency in the Shadows of the Fourth Industrial Revolution
This post is by Lilly Irani, Associate Professor of Communication & Science Studies at University of California, San Diego.
Amazon Mechanical Turk is but one platform in an ecology of automation and artificial intelligence. After more than a decade of operation, it offers crucial lessons in what is at stake when we talk about the “Fourth Industrial Revolution.” What’s at Stake in a Fourth Industrial Revolution? This essay works from the ground of AI practice and value chains to open up cautions, questions, and possibilities a world of production in which high-tech industries and their work technologies play a particularly central role.
The World Economic Forum writes that the Fourth Industrial Revolution represents the emergent forms of productive organization generated by “the possibilities of billions of people connected by mobile devices, with unprecedented processing power, storage capacity, and access to knowledge, are unlimited.” (Source) This revolution advances complex technologies, including artificial intelligence, robotics, autonomous vehicles.
In 2006, Jeff Bezos launched a new form of computer technology. Amazon had used the technology as a form of “artificial artificial intelligence” — data processing that could classify images, sounds, and texts automatically while still seizing on cultural nuances like humor, sexuality, and linguistic dialects. The service was part of Amazon Web Services, marketed alongside S3 and EC2 — just-in-time server space and computational cycles available to programmers through routine acts of coding. Bezos explained the third technology — the artificial artificial intelligence — as “humans-as-a-service.” That service was Amazon Mechanical Turk.
The secret of Amazon Mechanical Turk (AMT) was not achievements in computer engineering, statistics, or algorithms. In fact, AMT was born out of the failures of artificial intelligence to meet the needs of internet companies seeking to expand the domain of data they could store, classify, and serve up online. Rather, AMT offered a marketplace online where workers with computers and internet connections all over the world could flexibly complete data processing tasks around the clock. Employers seeking fast data processing no longer had to hire workers or an outsourcing firm; they would not even have to meet those who worked for them, whether online or face-to-face. They simply could place their data processing tasks online, set a price for each task, and design algorithms to receive, validate, and integrate workers” processed data into computer systems. The system allowed for a kind of massively mediated microlabor — large volumes of small, independent tasks distributed to large groups of workers.
AMT workers teach robots how to see not a stream of pixels but objects in space. AMT workers teach AI how group diverse photographs of what a person knows is “the same” product. AMT workers train search engine algorithms to adapt to cultural and linguistic change. There are no artificial intelligences and autonomous robots that work without the labors of data processing workers to train, calibrate, and categorize the myriad forms of unstructured data we produce in computerized networks.
Excitement about AI and robotics has proliferated work for such data processing workers. A decade after Bezos’ announcement, there are over 100,000 unique workers on Amazon, and several thousand workers available for data processing work at any time (Difallah et al. 2018). The majority of this workforce is employed in the United States and recent studies demonstrate that the median wage of this workforce is approximately $2 per hour (Hara et al. 2018; see also Berg 2015:557). These workers are not unskilled; 45% of US AMT workers and 91% of Indian AMT workers hold a college or post-college degree. In an International Labor Organization survey, 45% of US AMT workers said the most important reason they work is to supplement income from other jobs (Berg 2015:552).
The Ongoing Care and Feeding of AI
Robots and artificial intelligence must categorize culture — as relationships, as spatial formations, and as semiotics — in order to operate. Space and culture change with human activity and natural processes. Even as engineers produce AI and robots trained to perform in their specific symbolic and spatial contexts, those contexts are subject to change. AMT workers do some of the ongoing care and feeding that keep AI calibrated to culture.
Take an example from Twitter. Search engine engineers at Twitter faced a problem. Their algorithms built indexes that could quickly retrieve tweet results, ordering the results by prioritizing more common usages of the word. A search for “binders” might bring up tweets about office supplies. However, highly public events can change the meaning of a word or even generate whole new queries algorithms have never seen before. During the 2012 presidential debates, candidate Mitt Romney told a national audience he had “binders full of women” in reference to resumes. As people turned to Twitter to comment in real time, the search algorithms had no time to learn what a relevant result for such a query might be. Twitter engineers turned to AMT workers to quickly verify and adjust algorithms responding to new kinds of queries.
When we see an algorithm operating competently, we might perform what historian Geof Bowker calls an “infrastructural inversion” to ask how information is held in alignment with the world (1994). Systems like AMT keep thousands of workers at the ready to absorb the constant labor of teaching and ongoingly calibrating algorithms to keep them in alignment with a changing world.
Hype and Immobilization
What we know and hear about automation and AI is itself shaped by the forms of financial investment that resource these sectors. By hiding the labor and rendering it manageable through computing code, human computation platforms like AMT have generated an industry claiming that the future of work resides in the programming powers of master engineers and the algorithms and robots they produce.
Investor valuation strategies incentivize companies to hide the kinds of labor AMT workers do to train and maintain automation. Microwork companies, for example, attract more generous investment terms when investors perceive them as technology companies rather than labor companies. At one industry panel, a crowdsourcing startup CEO discussed the question, “Am I a labor business or a SaaS [software-as-a-service] business?” In response, a venture capital (VC) investor responded, “SaaS has a higher multiplier in the market. I was hoping it was a technology company and not a labor company when I invested!” Multipliers are rule-of-thumb quantities appraisers of various sorts — VCs, banks, buyers — use to estimate the value of companies. Multipliers represent an attempt to guess at the relation between capital investment and market value, whether that value derives from profits, revenue, or future resale. To act as technology companies, microlabor companies must convince investors, first, that their labor force is of little risk and of little cost, and second, that their technology confers an advantage over other companies. Microlabor companies do this in part by foregrounding algorithmic techniques for managing AMT-like workers and demonstrating a reliable flow of replaceable workers. As companies promise the ability to expand their operations quickly, so do they fuel scaling valuations.
Such claims constitute a form of hype — the sorts of promissory statements that generate value, are hard to dispel, and may not be true but fall short of fraud (Sunder Rajan 2006:114–115). By keeping AMT-like workers invisible, Fourth Industrial Revolution narratives obscure the very real questions of working conditions and class stratification that will persist or even grow as automation proliferates. The hype of inevitable AI and total automation immobilizes AMT workers from demanding better wages or improvement to their work conditions. Like the threat of workers made available cheaply in other parts of the world, public assumptions about AI and automation immobilize grassroots, democratic participation in shaping the future of work. This hype and the broader narratives of novelty and disjuncture that often accompanies it can prematurely disarm existing ethical and justice frameworks by which citizens can make claims about technological futures.
Collective Forms of Invisible Labor that Make AMT Work
Platforms like AMT, Lyft, and Uber make headlines both for spectacular new forms of work organization and questionable labor practices. In the shadows of these platforms, workers must innovate, cooperate, and absorb stresses to compensate for the shortfalls of these systems.
Amazon provides very little by way of support, training, or conflict resolution for AMT workers, perhaps for fear of appearing to be an employer rather than a procurer of services. In part to compensate for this lack of human resources infrastructures, tens of thousands of workers congregate on two major worker-run web forums; there, workers share advice with one another, negotiate norms of work, and struggle to establish more interactive and participatory relationships with employers (Martin et al. 2014). These collectives are sites where workers manage themselves and other workers, set norms, help employers, and sometimes coordinate work refusals. They advise employers about flawed task designs or bugs on forums. They teach each other how to use tools. These sites include TurkerNation, MTurkGrind, and the mturk and HITsWorthTurkingFor “subreddits,” among others. Turkers operate the forums, fundraise hosting feeds, and moderate communities. They regulate conversations, recruit requesters into collaboration, and educate newcomers. AMT workers’ are not compensated for these forms of invisible management consulting, training, and tool building. These fora are not unique to AMT. Scholars of Uber, Lyft, Didi, and Deliveroo similarly observe that workers create tools and forums that enhance their capacities to work and to gain a modicum of control over their work (Chen 2017; Rosenblat & Stark 2016).
There are many more questions we might ask about AMT a decade after its founding. For those long-term AMT workers who complete the bulk of the work (Berg 2015), what are long term effects of the work? Though AMT is formally flexible, workers often stay at their computers waiting for the posting of good work (Gupta et al. 2014). Anecdotally, forum moderators often struggle to raise funds for their time or server expenses from underpaid workforces. The labor of operating these forums and dealing with complex online relationships has led some workers to burnout. What are psychological and social costs of work that is flexible for employers but keeps workers on call? What kinds of repetitive stress injuries or occupational hazards accompany the extended and uncompensated labor of locating work and sustaining communities of mutual aid? How do firms’ need to emphasize the replaceability of workforces condition their production of public knowledge about technologies? How do these dynamics suggest transformations in disciplinary methodologies for approaching the study of technology.
Berg, Janine. 2015. “Income Security in the On-Demand Economy: Findings and Policy Lessons from a Survey of Crowdworkers.” Comp. Lab. L. & Pol’y J. 37: 543.
Bowker, G. 1994. “Information Mythology: The World of/as Information.” Information Acumen: The Understanding and Use of Knowledge in Modern Business, London: Routledge, 231Ð47.
Chen, Julie Yujie. 2017. “Thrown under the Bus and Outrunning It! The Logic of Didi and Taxi Drivers’ Labour and Activism in the on-Demand Economy.” New Media & Society, September, 1461444817729149. https://doi.org/10.1177/1461444817729149.
Difallah, Djellel, Elena Filatova, and Panos Ipeirotis. 2018. “Demographics and Dynamics of Mechanical Turk Workers.” In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, 135–143. WSDM ’18. New York, NY, USA: ACM. https://doi.org/10.1145/3159652.3159661.
Gupta, Neha, David Martin, Benjamin V. Hanrahan, and Jacki O’Neill. 2014. “Turk-Life in India.” In Proceedings of the 18th International Conference on Supporting Group Work, 1–11. ACM. http://dl.acm.org/citation.cfm?id=2660403.
David Martin, Benjamin V. Hanrahan, Jacki O’Neill, and Neha Gupta. 2014. Being a Turker. Proceedings of the 17th ACM Conference on Computer Supported Cooperative Work & Social Computing, 224–235. http://doi.org/10.1145/2531602.2531663
Rosenblat, Alex, and Luke Stark. 2016. “Algorithmic Labor and Information Asymmetries: A Case Study of Uber’s Drivers.” SSRN Scholarly Paper ID 2686227. Rochester, NY: Social Science Research Network. https://papers.ssrn.com/abstract=2686227.
Sunder Rajan, Kaushik. 2006. Biocapital. Durham,
 This case is drawn from ethnographic fieldwork and material from CrowdConf archives.
 Personal conversation with Kathleen Connolly on May 6, 2018, San Diego, CA.