Mechanical Turk Is Shutting Down, and the Robots Are Already Filming Your Chores

Amazon is closing down Mechanical Turk at the end of September. If you’ve never heard of it, Turk was the internet’s odd-job board: tiny, tedious tasks (tagging images, transcribing audio, answering surveys) farmed out to a global pool of humans for a few cents a pop. It’s been running since 2005, longer than the iPhone, and it quietly became one of the engines behind the AI boom. Somebody had to label all that training data. Turns out it was often someone in Manila or regional India or, honestly, someone in outer suburban Australia topping up a pension, clicking through image tags at 11pm.

The reason it’s closing isn’t a mystery. Recent reporting says something like 46% of the tasks on the platform were already being completed by AI, meaning the humans hired to train the machines were quietly outsourcing their own jobs back to the machines. There’s a joke in there somewhere, but it’s not a very funny one.

What struck me reading through the discussion wasn’t the Turk news on its own. It was seeing it sit right next to another story on the same forum: Figure AI, the humanoid robot company, has released a dataset of sixteen million videos of people doing ordinary household tasks, and they’re paying regular people to film themselves doing chores so the robots can learn to do them too. Someone in the comments mentioned people in India strapping a fifty-dollar phone to their forehead and cleaning house for a few dollars an hour, footage destined to become training data for a robot that will, eventually, replace exactly that kind of work.

So the pattern goes: first we paid humans cents to label the data that trained the models that write and code and answer emails. Now we’re paying humans a few dollars to film themselves vacuuming, so the models can learn to vacuum. It’s the same trade twice, just moving from the cognitive layer to the physical one. And the wage keeps sliding towards whatever the cheapest labour market on Earth will accept, because the whole point of the exercise is to make that labour permanently unnecessary afterwards.

I don’t think this is some grand conspiracy. Nobody sat in a room and planned “let’s get poor people to train their own obsolescence.” It’s just what happens when you optimise for cost and scale without anybody being responsible for the transition. The market doesn’t have a conscience, it has a spreadsheet, and the spreadsheet says human labelling is expensive and Claude can do it for less.

I work in tech, DevOps mostly, and I’ve had my own version of this feeling. Watching a script I wrote automate away a task somebody used to do manually gives you a small, complicated buzz. It’s satisfying to solve the problem. It’s less satisfying to think about the person whose Tuesday afternoon that used to fill. I’ve never had to make anyone redundant with a piece of code, but I’ve come close enough to understand the particular blankness of it: no malice anywhere in the chain, just a long series of reasonable decisions that add up to someone losing their income.

The bit that gets me about the Turk situation specifically is how invisible that work always was. Nobody put “trained the AI that eventually did my job” on a resume. It wasn’t glamorous, it wasn’t unionised, it barely counted as a job in most people’s minds, more like a side hustle for people killing time on the bus. But it was real money for real people, in places where a few dollars an hour is not nothing. When it disappears, there won’t be a press conference. It’ll just quietly stop mattering, the way the fax machine did, except the fax machine didn’t used to pay someone’s rent.

I don’t have a tidy solution here, and I’m suspicious of anyone who claims they do. Universal basic income gets floated every time this topic comes up online, and it might genuinely be part of the answer, but it’s also become a bit of a thought-terminating cliché, the “thoughts and prayers” of automation discourse. Said, nodded at, rarely funded. What I do think is worth sitting with is the plain fact that we are, right now, in the middle of the most significant labour transition since the industrial revolution, and most of the institutions meant to manage that kind of change (governments, unions, retraining schemes) are still moving at the speed of a council meeting about bike lanes.

None of which means I think the robots doing the dishes is a bad thing. If a machine wants to fold my washing, it can have the job, gladly. The tension is real and I’m not going to pretend it isn’t: I’m fascinated by what these systems can do and genuinely uneasy about who bears the cost of getting there. Both of those things are true at once, and I don’t think you need to resolve them to keep paying attention.

The Turk workers will find other things to do, some of them anyway, because people generally do. But it would be nice, just once, if the people building the systems that replace a job spent half as much energy on the people in that job as they do on the benchmark numbers. That’s not radical. It’s just decent.