How One Group of Amazon Workers May Be Making Themselves Obsolete
Waves of outsourcing have been killing manufacturing and call center jobs here in the U.S. for decades. Now a class of lower-paid tech workers may be working themselves right out of a job.
In 2005, Amazon launched the Mechanical Turk Service, a service marketplace where humans perform microtasks that computers don’t do as well as humans...yet.
Microtasks are small units of work that require human decision-making—things like identifying objects in photos, writing short descriptions, translating text from one language to another, and identifying emotion in written text. With 500,000 Turkers, as they are known, in 190 countries, some willing to work for as little as $1.38 an hour, they represent a huge informal workforce.
Many large-scale web companies use this collective power of human intelligence to make judgments, tag content, provide keywords, improve search results and reduce duplication of web pages, but there’s an added benefit when the collective decisions human workers make also contribute to machine learning.
Data scientist Jeremy Howard explains in his recent TED talk, “The Wonderful and Terrifying Implications of Computers That Can Learn,” “programmers once needed to program computers to do things by giving them explicit commands in excruciating detail, now with deep learning, a single algorithm can learn to do almost anything and gets better at it the more data you give it.”
Researchers have found that tapping into the large pool of Turkers provides them with subjects who can produce the large data sets needed in order to conduct experiments and develop algorithms, at a low cost.
Because each individual task may pay only 5 or 10 cents and can take anywhere from a few seconds to a few minutes to perform, it’s challenging to earn a living this way. Turkers are not always motivated by the money. They enjoy the flexibility of working from home, choosing their own hours, and getting paid right away for doing acceptable work. According to Amazon, many are “casual workers who work a few hours a week, for instance college students who work to buy books, stay at home mothers who work while the kids are sleeping and use the money they earn to buy groceries. Others are full timers who earn a living from Amazon Mechanical Turk. We also have casual workers who use Mechanical Turk for fun—they consider it an alternative to playing games online or watching TV. “
Mechanical Turk derives its name from an actual 18th-century automaton that played chess, duping opponents into believing they were playing a machine when in fact they were playing a human. Amazon’s clever name and tagline,“Artificial Artificial Intelligence,” points to the ruse of the human inside the machine.
Ironically, Howard, in his TEDTalk, cites one of the first examples of “machine learning” developed by Arthur Samuel who in 1956 actually programmed a computer to learn to play checkers. He did this by having the machine play checkers with itself thousands of times in order to “learn.”
Amazon's service is only one of more than a dozen microtask sites which include CrowdFlower, TaskRabbit, Fiverr, CrowdSource, Appen, CloudCrowd, and Microtask, all of which give companies and researchers access to low-cost human decision-makers. Much of this work is content management, content creation, research, translation and transcription and historically might have been done by entry-level production, editorial, marketing, or administrative assistants or maybe temp workers from an agency.
These were once good entry-level jobs that might lead to careers and knowledge, where a person gained understanding about the business they were working in and moved up the corporate ladder. Now, there's no ladder. While the workers who do this work may not mind, the question is whether companies will outsource more and more, cutting their regular workforce, and further destabilizing the job market.
CrowdFlower explains the breakdown is fairly evenly split as to why workers will work for pennies per task: earning money (29%), passing time (28%) and enjoyment (26%).
Because microtasking workers complete the jobs as independent contractors, they are not covered by any of the hard-won protections which American workers are legally entitled to: minimum wage, worker’s compensation and mandatory breaks, for starters.
In addition to lowering the cost of labor, these half-million workers also provide valuable data that researchers and companies can compile and use to create algorithms. In some cases, the algorithms train machines to do similar work.
Amazon uses the Mechanical Turk service to clean its database of over 500 million products, as well as to improve its search technology. While there’s no evidence that Amazon is developing algorithms to replace workers, other researchers have been able to use the marketplace to create large sets of data that in turn train algorithms to do the same type of work as humans.
In 2011, then-UC Berkeley researcher Subhranhsu Maji set up an experiment in which he used Turkers, as Amazon marketplace workers are known, to create 250,000 annotations on images—tracing outlines of a cat or a sheep, or answering questions about portions of photographs showing arms, legs or torsos. According to the summary of the research published online, “these annotations were used to train algorithms" to recognize people in different poses, what they were wearing, and other objects in images. In effect, the work humans did helped train machines to approximate the decisions humans were making. While workers are paid for each microtask, one can argue that the real value they provide is their contribution to training the algorithm which can be used over and over again.
Computer scientist and author of Who Owns the Future? Jaron Lanier has referred to crowdsourcing—information aggregated from multiple users—as “an informal economy [which] can be a road to ruin when companies use human labor to erase the value of people."
While Mechanical Turk and other microtask services appear to be a respite for low-skilled workers in the information economy, the seeds of self-destruction are built right in. Companies can use the collective human intelligence generated from hundreds of thousands of instances of humans making decisions, to create algorithms that will teach machines to automate some of these same tasks, and underpay, or ultimately eliminate humans in the process. This effectively means some of these lower-paid workers are helping machines get smarter and potentially putting themselves out of work.
Online companies have been using algorithms to improve their services, from improving search results and retail buying recommendations, to creating translation tools and developing image detection software.
As algorithms become more sophisticated and accurate, we’re likely to see more algorithms with less involvement from people. Google’s “new algorithm” learned to play Atari video games better than humans, and has implications for improving its self-driving cars, an innovation that has the future potential to impact workers who drive for a living.
In a PBS NewsHour segment titled “What Role Does Human Touch Play in the Digital Age?” Jaron Lanier mentions an automatic translation tool that relies on ongoing input from hundreds of thousands of translations by people to create an algorithm that “learns” how to translate. Lanier underscores the fact that some big data companies stand to benefit from the value of the algorithms yet humans may not be adequately compensated for the value they’re creating in the crowd-sourced economy.
Sometimes workers are aware that they’re training machines. In a post on Gizmodo, Eric Limer describes his brief stint as a Turk and the eerie sense he had that the work he was doing would ultimately be used in machine learning. In one job he was asked to make facial expressions in his webcam "so computers could learn to recognize emotion," he said. Describing another task, he said, “my mission was to watch a video and role play as a human being.”
Ultimately, machine learning could result in algorithms that replace work that is distressing for humans to perform. Computer science researcher Lydia Chilton cites an interesting example of how an algorithm could do work humans might find emotionally challenging, such as classifying images as child pornography. Chilton says if we automate classification "we can more easily check all the images on the web to stamp out this crime."
While crowdsourced data has been used to develop and improve algorithms for decades, some Internet companies are also conducting research experiments, without their users' prior knowledge. Facebook researcher Adam D. I. Kramer issued a “public explanation” on his Facebook page after leading an experiment in "emotional contagion" designed to track whether people are affected by seeing more positive or more negative Facebook posts from friends in their news feeds. Apparently they are. The research indicates that when users' feeds contained more posts with more negative sentiments, those users then posted more negatively. The reverse is also true—positive sentiments are also contagious.
Crowdsourced data is a valuable resource to companies that want to know more about their users or squeeze value out of other peoples’ collective intelligence. More transparency around this and simpler ways for users to opt out of aggregate or anonymised data projects would give users more control over when and how their behavior, choices and decisions could be used to benefit corporations.
Right now if people want to keep their data private and prevent companies from using it, there are “ways to do that,” though the process can seem daunting.
In an interview, Lanier said that, "using big data to not pay large numbers of people who are contributing is a rising trend in our civilization, which is totally non-sustainable. Big data systems are useful. There should be more and more of them. If that's going to mean more and more people not being paid for their actual contributions, then we have a problem."
In today’s digital gold rush, microtask workers are getting paid, though possibly shortchanged. If left unchecked, the logical end of “training machines” is human obsolescence. At $1.38 an hour, we’re almost there.