The Signal · Column 064

Amazon Is Closing The Factory That Taught The Machines

Andrei van Noordt · 20 September 2026 · 2 min read

Amazon is shutting down Mechanical Turk on September 30, along with SageMaker Ground Truth and Amazon Augmented AI. That is the whole human data-collection layer, retired on one date. Bezos called it artificial artificial intelligence, and the joke was load-bearing: for twenty-one years, when a computer could not make a small judgment call, the platform found a person who could. Hundreds of thousands of them. Much of what they decided became the training data the current generation of models learned from.

The easy reading is that the models got good enough and no longer need the help. I think the opposite is closer to the truth. The tasks Mechanical Turk handled were the cheap ones, the judgments a stranger could make in nine seconds with no context. Those are exactly the judgments models now make for free. What remains is the expensive kind: the call that requires knowing this domain, this dataset, this regulator, this failure mode. You cannot crowdsource that. There was never a marketplace for it, which is why Amazon is not building one.

The work did not disappear. It moved from a marketplace you could buy by the task to a payroll you have to compete for.

So the supply of human judgment feeding AI systems has not shrunk. It has changed shape. It used to arrive as a variable cost you could meter by the task, and now it arrives as a person on your payroll who understands your problem well enough to tell the model when it is wrong. Every company running serious AI tooling is discovering the same thing at roughly the same time, which means they are all reaching for the same small group of people. The ones who can label, evaluate and correct in a specialist domain were never abundant. They are now load-bearing, and the market has noticed.

This is the part most hiring plans miss. The AI budget line and the headcount line are treated as opposites, one growing as the other shrinks, when the deployment actually creates a new and specific hiring need. Someone has to own evaluation. Someone has to define what correct looks like in your domain and keep checking. That role did not exist on most org charts three years ago and it is not junior, because the whole value of it is knowing what wrong looks like.

For hiring leaders, the practical read is simple. Before you scale an AI system, name the person who judges its output and make sure they know your domain cold, because the era of renting that judgment by the task closes on September 30.