Amazon will shut down Mechanical Turk on 30 September 2026, having stopped accepting new clients at the end of July. Launched in 2005, this microtask marketplace let humans carry out, for a few cents, what machines could not do. Jeff Bezos described it as "artificial artificial intelligence". That phrase has never aged so well.
We have devoted two articles to the invisible workers of AI. Here is the end of the ancestor of that system, and the reasons for its demise are instructive.
What the service was
The principle was simple and, at the time, ingenious. A client would post thousands of microtasks — identifying an object in a photo, transcribing a handwritten word, verifying an address — and workers around the world would complete them for a few cents each.
The name referred to an eighteenth-century chess-playing automaton that impressed Europe, and inside which a human was hidden. The reference was deliberate: behind the apparent automation, human labour.
The service fed a considerable share of the datasets that trained the first vision and language-processing systems. Without it, part of what we use today would not exist.
What killed it, and it is ironic
Two causes combine, and the second is delicious.
Moving upmarket. The need has changed in nature. The goal is no longer to have anyone label images, but to have specialists evaluate complex reasoning. That is what we described regarding data labelling valued at twenty billion: the segment that is exploding is that of experts — doctors, lawyers, engineers.
Workers were using AI. This is the remarkable point. Studies have shown that a significant share of people performing these tasks had them processed by language models. In other words: a platform meant to supply human judgement to machines was in fact delivering machine outputs.
This reversal undermines the very foundation of the service. If you can no longer guarantee that the answer comes from a human, you can no longer sell human judgement.
This problem goes far beyond this platform. It arises wherever you pay someone to produce a judgement: model response evaluations, user testing, online surveys, proofreading. We have documented the difficulty of distinguishing generated text, and that difficulty has a direct economic consequence: human work becomes harder to certify, and therefore harder to sell as such. This is a problem that is only just beginning.
What is lost, despite everything
It would be easy to welcome the end of a system that paid a few cents per task. The reality is more ambivalent.
For some people — in countries where this income mattered, or for whom conventional employment was out of reach — this platform represented real access to paid work. Its closure does not automatically redirect them to something better.
And the replacement by platforms targeting qualified experts means this market is closing in on those who already hold credentials. Data work was one of the few digital activities accessible without a degree. It no longer is.
What to take away
This closure marks a shift in how AI consumes human labour.
For twenty years, the model was one of volume: many people, poorly paid, for simple tasks. It is being replaced by one of expertise: fewer people, better paid, for specialised judgement.
That is an improvement for those who reach the second segment, and a closure for the others. As often, the same evolution produces two opposite effects depending on where you stand, and only the first is the subject of press releases.
There remains the central irony, and it is worth keeping in mind. A service built on the idea that humans would do what machines cannot do is stopping because humans were using machines on it. That is a fairly perfect loop, and it probably heralds others.