Microworkers are “Disempowered to a Degree Previously Unseen in Capitalist History”
“Microworkers” are the anonymous digital contract workers whose labor powers the tech giants’ artificial intelligence systems. They’re hyper-exploited — and, like all other workers under capitalism, will continue suffering until they can organize.
Amazon’s Mechanical Turk is one of several large microworking platforms. (Daniel Acker / Bloomberg via Getty Images)
- Interview by
- Alex N. Press
Across the United States, worker militancy is on the rise. A labor movement that has spent years on the back foot is beginning to look fiercer and rejuvenated. The efforts of the ten thousand John Deere workers who struck and won a remarkably strong contract in particular have shown the sheer power of organizing. But how do you organize when your workplace is a computer and your job consists of countless data tasks for which you’re paid a pittance?
That’s precisely the problem facing microworkers around the globe. In his new book Work Without the Worker: Labour in the Age of Platform Capitalism, Phil Jones explains the uphill battle these workers face and how their precarious form of employment may be facilitating its own demise.
Jones joined Alex Press on a recent episode of Jacobin’s Primer, a podcast about all things Amazon. The following transcript has been edited for clarity and length.
Let’s start by defining microwork.
Microwork is really just another word for “short data tasks.” The tasks are hosted on digital platforms, which act as intermediaries between contractors and workers. The platform takes a cut of every transaction.
Contractors are very often tech giants like Google, Facebook, Uber, Microsoft. These companies need workers to process data and train their artificial intelligence technology.
Tasks on a platform might include annotating images of different areas to direct autonomous drones, annotating images of faces to train facial recognition technology, or training chatbots to recognize different accents or emotions in people’s voices.
I suppose Silicon Valley hype tends to depict artificial intelligence as something entirely unprecedented. But the relationship between new technology and workers really isn’t so different from the relationship between earlier technologies and workers. In both cases, a machine does some of the tasks once done by a worker. But the worker is still required to watch over the machine and correct its functions. That’s similar to what the microworker does.
The majority of those who do this work are in the Global South. They could be refugees, prisoners, or people living in occupied territories. Conditions are pretty miserable on microwork sites, generally speaking. One study found that, on a few of the platforms, a majority of the tasks are paid at less than twenty cents a task. Another study, years back, found that the average wage on one platform was $2 an hour.
Because the tasks are so short and poorly paid, workers often have to work super fast to reach subsistence wages. In fact, what often happens is that workers spend more time hunting for tasks than actually doing paid work.
You mentioned that refugees and prisoners are the ones often doing these tasks. In your book, the first chapter opens with an extremely disturbing scene. I’m going to read from it:
A woman living in Kenya’s Dadaab, which is among the world’s largest refugee camps, wanders across the vast, dusty site to a central hut lined with computers. Like many others who have been brutally displaced and then warehoused at the margins of our global system, her days are spent toiling away for a new capitalist vanguard thousands of miles away in Silicon Valley. A day’s work might include labelling videos, transcribing audio, or showing algorithms how to identify various photos of cats.
Amid a drought of real employment, “clickwork” represents one of few formal options for Dadaab’s residents, though the work is volatile, arduous, and, when waged, paid by the piece.
So, these are incredibly marginalized people. They’re in a place where this is all they can get for work, it’s all that is on offer.
Exactly. After those paragraphs, I go on to discuss Finnish prisoners who — through the company Vanu — were outsourced short data tasks on Amazon Mechanical Turk. What the big platforms get out of this is a sizable workforce that is in no position to organize or resist bad conditions. The platforms essentially get all of the work without any of the troubles typically associated with a workforce.
Since this is a show about Amazon, we’re especially interested in Mechanical Turk. Can you talk about when it went live, what it was originally used for, and how many people are working on it?
Sure. Mechanical Turk is the most infamous microwork site, but it’s actually relatively small. One conservative estimate from 2019 concluded that a quarter million people had performed a task on the platform. That’s, of course, much larger than the number of workers who use the platform as their main source of income.
Clickworker, by comparison, has a user base of over 2 million. Another platform, Appen, has a user base of 1 million. So Mechanical Turk isn’t really that big — partly because it mostly contracts workers in the United States or India, and not many other countries.
Mechanical Turk started as a service only available to programmers at Amazon. This was back in 2001, during the halcyon days of the dot-com boom, when the internet was still ostensibly great and we didn’t have widespread data mining yet. Amazon developed Mechanical Turk to solve an internal problem they had. Basically, its algorithms were failing to recognize product listings. But Amazon soon realized that there was a growing demand for cheap digital labor.
In 2005, Mechanical Turk went public and became a prototype for other platforms like the aforementioned Appen and Clickworker. The name of the platform is actually quite useful in helping us understand the initial logic behind it. The original Mechanical Turk was an outlandish device created in the eighteenth century by a Hungarian inventor. It was designed to resemble a chess-playing automaton. But it was no such thing. Tucked away inside the device was a human chess master. The point was to make it look as though a machine was doing something that humans were doing.
Amazon’s platform is a cheeky nod to this history. Jeff Bezos describes the workers as “artificial artificial intelligence.” The point of the site is that workers appear to contract as machines. This is partly a marketing ploy — another element of Big Tech’s tendency to make everything look high-tech even when it isn’t. But there’s also the theory that it’s set up this way to protect the contractor companies. If you outsource labor and hide it on tech sites, no one will know that your artificial intelligence is actually human intelligence.
Right, companies say they’re using artificial intelligence, and we just believe them. But often, they’re referring to incredibly low-paid workers.
That’s the main function of these sites. Companies hide the workers almost entirely, both to uphold their reputations and prevent organizing. Those workers are totally alienated. They’re often so far down the supply chain that they can’t see the end product they’re working toward. It creates an opacity where workers are disempowered to a degree previously unseen in capitalist history.
Profiting from Crisis
There are disturbing depictions in your book of what that means. One that comes to mind is of a favela resident in Brazil unknowingly programming a drone to monitor his own neighborhood. They’re building the tools that are oppressing them and people like them around the world, and they don’t even know it, which is an upsetting way of thinking about what the future of the economy looks like.
Part of the book’s argument is that we’re moving in this direction, and that’s rooted in a larger argument about employment cycles since the 1970s, profit, and capitalism’s evolution. Can you explain how we got to the point where microwork is a growing sector of employment? Also, please explain your use of the term “subemployment.”
My argument is one that has been made across a number of books. It starts with the work of historian Robert Brenner. In the 1970s, American manufacturing entered a recalcitrant period of stagnation in which it still resides. This stagnation bled into Japan and Europe.
It’s fundamentally a problem of overcapacity. Industries were producing more than could be consumed. So, a lot of companies began looking for cheaper ways to produce their goods. Competition, as is always the case under capitalism, necessitated this.
Companies started to outsource their work to regions where labor was cheaper. That led to fewer manufacturing jobs in countries like the United States and the UK. The workers displaced by this shift were transferred into the service sector, where job growth and productivity gains are notoriously much slower than in manufacturing.
This precipitated a crisis of labor demand. It’s something Aaron Benanav describes in his recent book Automation and the Future of Work. Concurrent with the lack of labor demand, many communist and colonized countries were opening their labor markets. Proletarianization was taking place in those countries, which expanded the global supply of labor.
The upshot of this — counter to other theories — has not been apocalyptic levels of unemployment in the Global North, as is repeatedly predicted in the literature on automation. Instead, we’ve seen consistent downward pressure on wages, worse conditions, insufficient hours, and widespread market volatility.
A lot of the jobs created in recent years have barely differed from the most abject forms of joblessness. In the book, I describe this as “subemployment.” The prefix captures the idea that many people hover in a strange nether region between employment and unemployment.
Simultaneously, due to structural adjustments and deregulation (policies largely coerced by institutions like the World Bank and International Monetary Fund), the informal sectors in Global South countries began to grow. That became a space to absorb workers pushed out of public services and agricultural work. You end up with a global pool of surplus workers, which puts continual downward pressure on wages and conditions. It also stifled the possibility of organizing.
So the “sub” in “subemployment” in part refers to inadequate working conditions. One example that you lay out in the book is wage theft, which is taking place at every stage of labor on microwork sites.
Absolutely. Wage theft is rife on platforms like Mechanical Turk and Clickworker. One survey found that a third of workers on microwork sites experience wage theft regularly. On Clickworker, a whopping 15 percent of all tasks go unpaid.
This happens partly because workers are given no recourse to action on the platform. If a task is thought to have been done poorly and gets rejected by a contractor, workers can’t really do anything about that. Contractors can review workers, but not the other way around. So, if contractors don’t pay, there’s no way for a worker to flag that and let other workers know.
This allows wage theft to flourish and take many forms. One common way is for lots of workers to be given the same task. The workers who answer in the majority are paid, while the minority are assumed to have done a bad job and are denied payment.
Imagine that there are two photos. One is of a cat and the other is of something else. If a majority say that there are two cats, everyone who correctly identified that there is just one will go without payment.
Exactly. It could mean that the majority is wrong and they’re getting paid, while the minority is right and they’re not getting paid. It’s not even a precise way of rewarding good performance.
Another way we see workers getting less than promised, which is effectively wage theft, is that requesters could say a task takes much shorter than it actually does. They could say it takes fifteen minutes when it really takes forty.
On Mechanical Turk, for instance, time restrictions are only indicators of how long a task should take. So, the worker will see the task advertised on the platform and think, “Great — a dollar for fifteen minutes of work.” But because these time limits are defined by requesters who are eager to cut costs, a task might be marketed as $1 for fifteen minutes when it actually takes closer to thirty minutes. The worker might not realize this until they’re already fifteen minutes into the task. And if you back out at that stage, then you won’t be paid.
Wage Thieves
Everyone knows that bosses often understate how much work their employees do. It’s a common complaint: Your boss says something should take ten minutes when it really takes thirty. So, I’d imagine this is happening constantly for microworkers.
It must be. Even reports from relatively conservative institutions like the International Labour Organization suggest that wage theft is rampant.
The fact that wages on these sites are called rewards and tokens says it all. I argue in the book that microwork moves us from wage to wager. Paying workers has essentially become discretionary as opposed to contractual. As a worker, you don’t have any guarantee that you’ll be compensated for performing a task.
When I think of jobs that go unpaid, one thing that comes to mind are high-prestige internships, the idea being that those jobs will pay off in the future. But we’re talking about things that pay $1 for fifteen minutes of work. There’s no future payoff. This is just your source of income.
For lots of people who use these sites and don’t rely on them as a primary source of income, they probably experience that and just leave. It isn’t worth their time. But for a lot of people, this is their primary source of income. There are plenty of workers in India and Venezuela, for example, who are totally dependent on microwork. It’s basically their full-time job.
If those workers are continually not given the wages they were promised, that means a significant amount of stress. It introduces a level of uncertainty about where you’re going to be at the end of the week. There haven’t been any studies about the mental health of microworkers, but those would be really interesting to see.
You’ve mentioned Clickworker and Appen in addition to Amazon Mechanical Turk. In your book, you divide microwork platforms into two types: curated crowd and crude crowd. Can you just lay out this bigger picture? Who are the major players in this industry, where are they operating, and what are these two categories?
Crude crowd sites, as the name suggests, basically let any worker or contractor use the platform. It’s a crude form of employment. Any kind of computational task can be uploaded so long as it fits with the site’s functions.
These sites consistently pay below subsistence wages, according to the standards of pretty much every country across the globe. Crude crowd sites are the type often associated with the highest levels of wage theft. Included within this category are Amazon Mechanical Turk and Clickworker. They’re the two most notorious microwork platforms. They’ve appeared in endless stories in the New York Times, the Guardian, and so on.
Then there’s this other layer of microwork platforms which have received less coverage. That’s partly because the labor conditions are slightly less miserable on these platforms but also because they’ve been good at hiding behind multiple software architectures. Platforms like Appen and Playment cater almost exclusively to the machine-learning needs of large corporations. So, they might have workers processing data for warehouse robotics, chatbots, and so on. They tend to also host tasks that require significant skill. A worker might have to complete an unpaid test to gauge their competence before they can start a task.
The book mentions that microwork platforms can host translation tasks. Most people see translation as an esteemed skill that is difficult to acquire, but microworkers are translating for very small amounts of money.
This is part of a disturbing trend. White-collar professions are being proletarianized on these sites. Translator is a good example. What you’ll find on these sites are lots of short translation tasks, which are largely being used to train Google’s translation machines and similar software.
Instead of a contractor having to hire a professional translator with labor rights, a minimum wage, and access to a union, they get forty workers completing all those translation tasks — none of whom will have any of those basic protections. It creates a flexibility for capital that neoliberals before now could’ve only dreamed of. You can effectively hire and fire a workforce in the space of an hour. You don’t need any long-term employees, who carry the potential of causing trouble.
“Turkopticon”
You just laid out how microwork sites dismantled well-paid, white-collar professions and created a low-wage, easily fireable workforce. What is the effect of this on organizing? What are some possibilities for these workers?
With the exception of Mechanical Turk, there has been very little worker organization on these platforms thus far. Workers haven’t unionized mostly because they’re remote and can’t readily communicate with one another. They’re oppressed by platforms that are designed to disempower them.
Also, it’s very difficult for unions to see this workforce. They’re essentially invisible. These aren’t easy workers to contact. It’s not like you can send somebody into a microwork platform and organize workers.
Nonetheless, Mechanical Turk workers have set up a forum to pressure the company into providing better conditions — specifically around issues of wage theft and rejected tasks. This is something I heard about from a colleague. I’m excited to hear more in the coming weeks.
Mechanical Turk workers in the past have taken action against dodgy contractors. Earlier, I mentioned the review systems that only allow contractors to rate workers. Workers have pushed back against this one-sidedness by creating a tool called Turkopticon: a browser plug-in that overlays the screens of workers who download it and allows them to write reviews of contractors, which they can then upload in real time. Turkopticon allows other workers to see if there’s a dodgy contractor on the platform.
The problem, though, is that there have been attempts to do stuff like this in the past on Mechanical Turk, and they’ve gotten shut down very quickly by the platform — partly because they’ve relied on the platform itself to authenticate workers for the forum. It’s this difficult situation where microworkers are dependent on software architectures that have been created to prevent them from organizing.
In the book, I don’t offer too many organizational strategies because I feel like it should be left to the workers to figure this out. It feels somewhat condescending for me to come up with those ideas.
But I do work through a couple of potential scenarios that might help inform future strategies. For instance, data sabotage might be one way that microworkers could push for better wages and conditions. If enough workers agree to perform their tasks badly, clogging the flows of data on which these platforms rely, then you could see all kinds of chaos on the internet. Imagine if a bunch of Facebook moderators decided to withhold their labor for a couple of weeks. Facebook would probably have to shut down their platform because it would be swamped with pornographic and traumatic imagery.
Similarly, Google raters find work through platforms like Appen and Lionbridge. If they decided to withdraw their labor, Google searches would no longer be personalized in ways that are particularly useful. It would be pretty catastrophic for these companies.
In writing about labor, I’ve talked a lot about how to organize rideshare workers. Traditional unions have a hard time locating drivers; it’s a very difficult task. But since drivers are somewhat congregated in big cities, organizing can happen in parking lots, garages, or near the airport — places where drivers stop in between customers.
Microwork platforms don’t offer that. If this is a growing workforce, it’s a growing problem for the labor movement.
The worker center model could be useful here. These centers provide day laborers with support, food, and an opportunity to take a break and meet other workers. Something like that could be really helpful for microwork, particularly in cities across the world where there’s a high concentration of microworkers. That’s something researchers could get data on. They could figure out where it might be most effective to open these worker centers.
The last thing I want to talk to you about is the role that data is playing in this process. The data from every task that’s completed through Amazon Mechanical Turk, for example, stays with Amazon. The general understanding is that workers are useful because they produce profits. But in the book, you say that they’re producing data.
Ultimately, data leads to profits for these companies. Microwork platforms aren’t just providing a platform for contractors and workers to meet and then taking a cut from those transactions. If you look through the terms and conditions on some of these sites, they suggest that information about the labor is more important than the product of the labor.
The goal of microwork is to show an algorithm how to drive a car or to show a drone how to surveil an urban area. Workers are directly showing machines how to do their jobs, basically. So, the real role of microworkers could be to automate their jobs away.
This isn’t actually that speculative. Companies like Google and Facebook have been pretty explicit about their desire to automate content moderation tasks. If you take the time to read Amazon Mechanical Turk’s terms and conditions, they effectively say that the data from any task completed on the platform can be used by Amazon for its own machine-learning purposes.
Now, it would make sense that the reason why Amazon took the Mechanical Turk platform public — rather than keeping it as an internal service — was not really to take a cut from transactions between workers and contractors which, as I just pointed out, isn’t even that profitable. It’s more likely that they wanted to gain wider access to available data. From Amazon’s perspective, there’s always the possibility that, while hovering over the transactions that take place on Mechanical Turk, they might find something useful for their web-services ecology.