Field Notes: The Future of Work
The Machine Finally Stopped Needing Us to Pretend
Amazon is closing Mechanical Turk after twenty-one years. What replaces it says more about the next decade of GTM and data work than the shutdown itself does.
A recent study found that as many as 46 percent of the workers on Amazon's Mechanical Turk platform, the people paid, by the task, to do small pieces of digital labor that computers supposedly couldn't do, had started quietly using AI models to do that labor for them. Read that twice. The workers hired to fill the gap between human and machine judgment had started closing that gap from the inside, on their own initiative, for their own reasons, well before their employer noticed or cared.
This isn't really a story about Amazon, though Amazon is the one making it official. The company announced last week that it will shut down Mechanical Turk on September 30th, ending a twenty-one-year experiment that, at its peak, paid something like half a million people across nearly two hundred countries to label images, transcribe audio, and answer surveys for a few cents a task. The closure notice ran two sentences. There's no severance, no transition program, just a support page and five weeks to withdraw your balance. That part is almost beside the point. Companies retire products unceremoniously all the time. What's worth sitting with is the phrase Jeff Bezos used to describe the thing back when Amazon was proud of it: "artificial artificial intelligence." Humans, pretending, with Amazon's full institutional blessing, to be the machine intelligence that didn't exist yet.
The name itself carries more weight than people give it credit for. The original Mechanical Turk was an eighteenth-century chess-playing automaton, and it was a hoax: a cabinet with gears and a wooden mannequin, and inside it, invisible, a real chess master pulling levers. Audiences wanted so badly to believe in the machine that nobody thought to ask who was hiding in the box. Amazon's version inverted the joke. It put the human in plain sight and called the whole system "artificial" anyway, as if the labor only counted as real once you'd built enough abstraction on top of it to forget a person was doing it. That's not a small linguistic tic. That's the entire economic logic of the last two decades of digital outsourcing, stated out loud, by accident, in a product name.
It would be easy to tell this as a simple morality tale (humans replaced by robots, curtain falls), and that's not quite what the data shows. What's actually happening is stranger, and more revealing. The market isn't disappearing. It's splitting, hard, into two pieces that have almost nothing to do with each other anymore.
That's one half of what's replacing Mechanical Turk: a credentialed, well-capitalized market for expert human judgment, recruiting doctors and lawyers and software engineers to rank model outputs and catch the kind of subtle reasoning failures that no anonymous crowd could reliably catch. It is, by any measure, an enormous and fast-growing industry. It is also not a replacement for what Mechanical Turk actually did, because Mechanical Turk's real function was never expert judgment. It was the opposite: it was the lowest possible barrier to entry, the version of digital work you could do with no credential, no network, and no local opportunity worth mentioning, from anywhere with a laptop and a spare hour.
The other half of the split is the part that doesn't show up in the funding announcements. Bulk labeling, image tagging, transcription, the unglamorous multilingual work that still resists full automation, hasn't gone away, but it's consolidated into a smaller set of larger vendors operating more like traditional outsourcing firms than open marketplaces, still drawing heavily on labor in the Global South, still priced nowhere near the expert tier, and shrinking a little more every quarter as automated labeling pipelines pick off the easy cases. Reporting on AI-training contractors in Kenya has documented take-home pay as low as roughly a dollar thirty an hour for reviewing some of the most disturbing text on the internet, so that the resulting models would learn not to produce it. That work sits adjacent to Mechanical Turk rather than on it, historically, but it's built on the identical premise: that the cheapest available human judgment, sourced globally and priced in cents, is an acceptable input for building the next generation of software.
Here's the thing worth arguing plainly, even if it's a little uncomfortable: the rise of the expert tier does not fix the problem with the bulk tier. It just makes the bulk tier easier to stop thinking about, because now there's a shinier, better-paid, more defensible version of "humans training AI" to point to in the press release. Two extremes are coexisting inside the same industry, and only one of them is getting the funding rounds.
A broader version of this same pattern isn't unique to data labeling. It's the same logic reshaping outsourced call-center work, first-pass content moderation, and, in a much smaller, more mundane way, the browser extensions a sales rep used to click all day to look up a lead's email address. In every one of these cases, an entire layer of the economy got built on the premise that human judgment, dispersed globally and priced as cheaply as possible, was the most efficient available input for a task. That premise held for a long time because it was true. It is now false for a rapidly widening set of tasks, and every market built on top of the old premise is operating on borrowed time, whether or not anyone has gotten around to posting the two-sentence notice yet.
None of this deserves a tidy ending, in either direction. It is not simply progress, because the version of low-barrier income that Mechanical Turk offered (messy, underpaid, entirely unprotected as it was) genuinely doesn't have a replacement in the market that's emerging to take its place. And it is not simply loss, because the thing it's being replaced by is, in places, a real improvement: better pay, more professional standards, actual leverage for the workers who can clear the new, much higher bar to entry. Both of those things are true, for different people, at the same time. That tension, who gets the new bar and who gets left below the old floor, is going to be the actual policy fight, once anyone in Washington notices there's one to have.
The name always gave away the joke. A machine that looked intelligent because a person was quietly doing the thinking inside it. Mechanical Turk is closing not because that joke stopped being true, but because it finally stopped being necessary. The thinking inside the box is, more and more, actually being done by the box. What replaces it for the people who used to be the ones hiding inside it is still an open question. Nobody has announced a plan for that yet. That's the part of this story that doesn't end on September 30th.
Where LeadGenius fits
We started as an answer to this exact problem.
LeadGenius began in 2011 as MobileWorks, one of the earliest alternatives to Mechanical-Turk-style crowdsourced task work. We built our model around a different bet: pair AI with a human-in-the-loop research network, and pay that network according to local cost of living rather than race-to-the-bottom task pricing. It's the same tension this piece raises, applied to how we actually source data.

.avif)

