Big Tech’s Algorithms Are Built With Invisible Labor
Don’t let the shiny products and ping-pong tables at work fool you: tech platforms are just as reliant on a wretched, toiling workforce as any other company under capitalism.
A man works from his home in Manila, Philippines, 2020. (Lisa Marie David / NurPhoto via Getty Images)
Today we are so habituated to online platforms that it’s common to refer to trying to influence “the algorithm.” From what appears in your Twitter or Instagram feeds to what Google proposes in the search bar as you type, to the ads you’re served at every turn, hidden algorithmic processes serve you content and learn from your selections to “train” the algorithm to better meet your needs next time.
But while these unseeable processes shape our online experience, labor is exploited behind the scenes. For an enormous amount of algorithm development, especially in today’s fast-moving AI industry, much of the training is done by “microworkers” — men and women, often in the developing world, who are training these systems manually.
For literal pennies, they take on tasks that may last for minutes or less, like flagging violent social media posts or ranking the relevance of search results. The workers performing these “human intelligence tasks,” or HITs, are as essential for the gigantic platform companies as they are utterly exploited and abjectly disorganized. Phil Jones’s book Work Without the Worker: Labour in the Age of Platform Capitalism allows the reader to see these invisible hands and the challenging frontier of worker organizing they represent.
While the mystique of tech in the United States has made most people content to believe algorithms train themselves to find our dinner reservations and sort our pictures, Jones observes the reality is that “the magic of machine learning is the grind of data labelling. Behind the cargo cult rituals of Silicon Valley is the grueling labour of sifting hate speech, annotating images and showing algorithms how to spot a cat.” Thousands of these simple HITs keep our user experiences smooth, and far more are needed for still-fanciful new technologies like the elusive self-driving car, training the alleged autopilots how to tell a street sign from a pedestrian.
These twenty-first-century proles log into Mechanical Turk, Appen, Clickworker, or other microwork platforms to complete tasks that often take just minutes or even seconds. Task requesters have near-total advantage, as they have no visible attached ratings from past task-completers; workers, meanwhile, are rated.
Given the tiny nature of the jobs, the workers seldom know the full purpose of the work they do or the identity of the requester, creating an enormous information asymmetry. The platforms encourage more workers to sign up than can be supported by the existing tasks, keeping productivity up and compensation down. The lack of physical workplaces, or indeed any recognition of fellow workers beyond an online avatar, tilt the landscape further against the workforce, whose accounts are deletable by the platform at any moment.
Use of these platforms is enormous in surprising places, like translation, where its use has exploded. Jones observes that the use of penniless microworkers to translate driving directions and appliance instructions “displays in stark terms capital’s violent path through the ‘vocations,’ turning professionals into proletariats.” But Jones examines the larger claim that gigantic proportions of jobs are about to be automated. He observes that rather than the common fear of “erasure of whole jobs,” the more common effect is “adaptations to a given job’s task composition and . . . the overall quality of the work.” Many workers will recognize that new technology in their jobs has meant continued employment with new tracking and fewer enjoyable aspects of the job.
Various forms of “subwork,” from Uber contractors to welfare recipients forced into involuntary “workfare,” are the more likely outcome of AI’s near-term role in capitalism’s evolution. With his trenchant wit, Jones brutally observes that “automation doomsayers . . . forget that a particular technology is generalized only if it turns out to be cheaper than employing a workforce.”
Mechanical Jerks
Some of the larger microwork platforms have clients that include the very biggest of Big Tech, including Google, Microsoft, and Amazon (owner of Mechanical Turk). The attraction of the microwork contractors is obvious — Jeff Bezos himself called Turkers “artificial artificial intelligence,” recognizing their role in making today’s search and content-moderating algorithms look far more automated and efficient than the reality.
Google hired the large microwork platform Appen (previously called Figure Eight) during its work for Project Maven, a Defense Department contract for an AI program to sort drone video footage. Much as with “autonomous vehicles,” AI drone targeting requires enormous amounts of data sorting and labelling for the program to itself “learn” to discriminate among its giant flow of continuous camera data. Taskers, of course unknowingly, did CAPTCHA-like selections from images to discern buildings from hills and cars from people. Appen also contracts with Microsoft and Amazon Web Services on various projects, while Twitter uses Turk to swiftly identify trending queries.
Amazon is also suspected of using Turk to train its alleged facial recognition software, sold to police departments nationwide. And Google created its own microwork platform, Raterhub, for the crucial task of testing and maintaining its search result quality (other than stuffing the results with paid ads). Contract search raters, primarily from the Philippines, rank the quality and relevance of search results and flag offensive or illegal items. Google’s trillion-dollar search algorithm is kept on its toes by workers in a country with an average income of $3,430 in 2020.
For all the talk of how microwork can provide development opportunities for the global poorest (much like its spiritual predecessor, microloans), the reality is different: the average Turker makes a wage south of an abominable $2 an hour. Turned to by international refugees, developing-world slum residents, and developed-world sub-employed postindustrial workers, the idea of microwork as a career or development strategy is laughable.
Jones reviews the formidable obstacles to organizing workers who have little to no user profiles, no messaging system, and no physical meeting place. Worse, they are workers who undermine their own employment, because microwork platforms like Mechanical Turk record all possible data regarding completion of requested tasks — how long it took, the steps taken, and their order — in order to ultimately increase the level of automation and show the algorithm “how to fulfill the role of labour.” This leads Jones to observe that while Turk “may appear as a labour broker . . . its real purpose is to provide data to Amazon Web Services.”
With their utter lack of potential avenues for worker solidarity and the near-total advantage of the anonymous task requester, Jones concludes that “microwork represents the apex of neoliberal fantasy: a capitalism without unions, worker culture and institutions — indeed, one without a worker capable of troubling capital at all.” Jones’s rich, cultivated contempt for these platforms yields a rich prose that makes the book easygoing, as in this description of a typical microwork setting: “Cramped and airless workspaces, festooned with a jumble of cables and loose wires, are the built antithesis to the near celestial campuses where the new masters of the universe reside.”
The reliance of the platforms on the microworkforce is real enough that Jones notes that “a strike by content moderators would instantly swamp user feeds with violent and pornographic images.” But capital has not failed to foresee this. When Facebook noncontract workers walked out over then-president Donald Trump’s use of the platform to spread heinous lies, a moderator statement read in part: “We would walk out with you — if Facebook would allow it. As outsourced contractors, non-disclosure agreements deter us from speaking openly about what we do and witness for most of our waking hours.”
Despite it all, Jones uses lessons from past organizing successes to recommend a few strategies. Rather than full-on organizing for collective bargaining, he describes actions similar to the 2018-19 Uber strikes in Cape Town and Mumbai, which blocked central arteries over demands for cheaper fuel, and actions by drivers in London and Paris over being made to carry the burden of climate policies. He also makes a strong case for centers for wageless workers to congregate, wageless worker associations, and demands to de-commodify goods like health care. More playful tactics are also explored, including Turkopticon, a browser plug-in that overlays the site interface and allows real-time reviews of the requesters. He concludes that for these workers, “the best chance of resistance lies in forging wider alliances.”
For all their gigantic data centers and smooth, opaque interfaces, at bottom, tech platforms are just as reliant on a toiling workforce as any other company under capitalism. Yet the utterly opaque nature of the online economy is such that users have little to no idea of this reality, and the cult of tech helps the platforms insist we pay no attention to the man behind the curtain. In its beautifully fiery condemnation of global exploitation, Work Without the Worker makes a reader hanker for a future world of capital without the capitalists.