What the End of Amazon’s MTurk Tells Us About Today’s AI Race
Human labor fueled the rise of AI. It can also be used to halt it.
Krista Pawloski
I’ll never forget my first job working on Amazon’s Mechanical Turk project, back in 2008. I was on maternity leave and needed some extra cash to make ends meet, so for several hours a day I would look at and tag photos with as many keywords as possible. If I was looking at a picture of a cloud in the sky, for instance, I would tag additional words such as “sky,” “daylight” and “atmosphere.”
I didn’t know then that Mechanical Turk, or MTurk as it came to be called, would become my fulltime job and that in the process I would become one of the hundreds of thousands of data workers who built the backbone of Big Tech’s AI machine.
After a little more than two decades of service, Amazon shut down MTurk for good on September 30 of this year. For me and my fellow “Turkers” who worked on the project, the moment was bittersweet, but not unexpected — Amazon has been quietly divesting for years, allowing competitors like Outlier, Appen, Mercor, Telus and Prolific to take a bigger part of the market share.
The end has been more of a whimper than bang, especially compared to the waves MTurk made when it first came online. At the time of its inception in 2005, Mturk promised to revolutionize work and provide clients — many of them Big Tech companies, universities, and government agencies — with what Amazon CEO Jeff Bezos called “artificial artificial intelligence.”
TechCrunch heralded it as “brilliant” in 2005 with the headline, “Amazon finally relevals itself as the Matrix.” It pointed to the very genesis of MTurk, what Venky Harianaryan, the Amazon manager behind MTurk, called a “hybrid machine/ human computing arrangement.”
MTurk was at the cutting edge of computing and research in the early aughts, but from the very beginning the reality did not match its shiny veneer. The human intelligence aspect of MTurks’ machine-human hybrid was completely devalued, with many of us working for pennies on the dollar for many jobs, all while training the algorithm that would one day replace us.
Named after Wolfgang von Kempelen’s 18th century Mechanical Turk — a machine that could supposedly play chess autonomously, but was actually operated by a human hidden inside the machine — Amazon also tried very hard to obscure the network of human labor that made MTurk possible. Turkers were supposed to be anonymous, unknown to the requester and each other, in order to keep our labor cheap and unorganized. And for a while it worked, because MTurk was like the Wild West.
The clients, or requesters as we called them, (and which used pseudonyms) would often pay far below minimum wage. There was also no oversight, so the requesters could deny a Turker had done the job and refuse payment with no repercussions. Turkers could spend hours fighting requestors or start a project that was falsely advertised in terms of time commitment or content. According to a 2016 Pew Study, carried out at the height of MTurk gig work, more than half of Turkers made less than the federal minimum wage: $7.25 per hour.
When I first started turking, I took a number of low paying jobs, but as I started to rely on MTurk for full time work, I became more savvy and started looking for help from other Turkers, because while Amazon wanted to keep us anonymous and apart, we still found each other. We connected on Reddit message boards, Facebook, Twitter (now X) and other online forums, passing along tips on who was a bad or a good requester.
In 2008, student researchers Lily Irani and Six Silberman created Turkopticon as a class project. It started as a simple way to quickly rate requesters so Turkers could weed out the bad jobs and protect each other, but over the years it grew into a worker-led organization that united Turkers around the world. I became one of the lead organizers of Turkopticon in 2020.
Over the past few years, we’ve helped over 1,000 Turkers fix erroneous suspensions and maintained a database of more than half a million requester reviews. And as more data workers moved onto competing platforms, we’ve extended our support beyond MTurk, while never losing focus. Until the very end, we supported Turkers with payment issues, advocated for platform improvements and got workers reinstated after Amazon kicked them off the platform with no explanation, something that happened periodically every few months.
When news broke that Mturk was planning to shutter for good, dozens of Turkers reached out to me. Looking back now, it’s clear to us that most of the jobs we did on MTurk helped Big Tech companies train the models and algorithms that now power so much of our AI ecosystem. Requesters may not have used their company names, but after saying “Hey Google” 72 times into a microphone for a day, I am pretty sure I know who I was working for.
We laughed remembering our favorite jobs — one of my personal favorites was helping train AI to identify brick kilns in satellite images to help track child labor violations — and reflected on the harder content moderation jobs that exposed many of us to hundreds of disturbing images, including beheadings, to train Youtube’s AI to moderate the platform’s videos.
Bezos may have called us “artificial artificial intelligence,” but the reality is there is nothing “artificial” about AI, because all along it was people like me that made it all possible. It’s our talent and our labor that’s fueled the current gains in AI. We taught the models what a human voice sounds like, what’s right, what’s wrong, and how to interpret the world — and because we made it, we have the power to change it.
Today, more than half of Americans are wary of AI with many citing job loss as a major motivating fear. It’s a fear that is now a reality for myself and the other Turkers. We trained the machines, and just as these machines become more and more advanced, MTurk has shut down. The corporations driving the AI race, including Amazon, want us to believe that the AI world they are building out is inevitable and out of our control, but that’s not true.
We were just supposed to be atomized workers, but we organized to build something stronger and better. And while MTurk may no longer exist, the work we did and the lessons we learned together can serve as a guiding light forward: we are stronger together and just like we collectively built AI, we can also stop it, too.
Thank you for reading this story. At a time when corporate media is dominating the web and our airwaves, independent journalism is more important than ever, and that’s what we provide at In These Times. We offer some of the best reporting, investigations and analysis on the growing movements for justice in the U.S.—and our work is firmly pro-labor and anti-fascist. To continue producing it, we need your support.
If you believe in the power of an independent press, please make a donation to In These Times today. By becoming a sustaining member, you can ensure that our work reaches even more readers.
And if you become a sustainer right now, you'll get a free year-long print subscription and a special subscribers-only print from artist Molly Crabapple.
Thank you.
Get 10 print issues of our award-winning, worker-first journalism here!
Krista Pawloski is a data worker and Turkopitcon organizer.
