The Ghost Workers Who Trained AI Are Being Shown the Door

The chatbots and image generators you use every day did not appear out of thin air. Behind them sits two decades of invisible human labor: millions of people tagging photos, sorting sentences, and grading AI answers for pennies a task. The platform that came to symbolize that work, Amazon Mechanical Turk, has quietly stopped accepting new customers. The hand that raised AI is now being pushed aside by it — and the irony is hard to miss.

What Mechanical Turk Actually Was

Amazon launched Mechanical Turk in 2005, and even the name tells a story. In the 18th century, a chess-playing automaton called “The Turk” astonished European courts — until people discovered a human operator hidden inside, moving the pieces. A machine on the outside, a person on the inside. Amazon borrowed the name on purpose and called the whole thing artificial artificial intelligence.

The mechanics were simple. Companies broke down small tasks that computers struggled with — so-called HITs, or Human Intelligence Tasks — and posted them by the thousands. Workers around the world picked them up for a few cents to a few dimes each: identify the object in this photo, flag this spam comment, answer this survey. Boring, granular, endless.

The Real Foundation of the AI Boom

Here is the part that matters. The modern AI industry was built on top of this platform. To train a model that recognizes images, someone first has to attach labels to millions of photos — this one is a cat, that one is a traffic light. The place that did that labeling at scale, and on the cheap, was Mechanical Turk.

ImageNet, the dataset that arguably kicked off the deep learning era, was assembled this way. More than 14 million images were tagged by hand, and a large share of that work passed through Mechanical Turk workers. When the chatbot era arrived, their role did not end — it shifted. Reinforcement learning from human feedback (RLHF), the process of having people pick which of an AI’s answers is better, leaned heavily on the same kind of crowd labor. The reason today’s assistants sound polite and useful is, in part, that thousands of anonymous workers ranked their responses one pair at a time.

The Shadow Called Ghost Work

There was always a darker side. In their 2019 book, anthropologist Mary Gray and computer scientist Siddharth Suri named this kind of labor ghost work. The output shows up as a sleek, seamless AI product. The people who made it stay firmly behind the screen.

The problem was how they were treated. Many workers earned well below minimum wage per hour. Whether a task got approved was decided unilaterally by the company that posted it, and a rejection — even an arbitrary one — came with no real avenue for appeal. No employment protections, no benefits, just pay-per-task piecework. For years, the glossy narrative of an AI revolution sat on top of this low-wage labor, and almost nobody talked about it.

Why the Doors Are Closing Now

So why now? The uncomfortable answer is AI itself.

Tasks that once required a human — labeling, sorting, classifying — can increasingly be handled by AI models directly. Using one model to prepare the training data for the next is now faster and cheaper than paying a crowd of people to do it. AI, trained on human labor, is now quietly removing the demand for that labor. The material that built the machine is being consumed by the machine.

The market shifted underneath Mechanical Turk in another way too. Specialized data-labeling firms like Scale AI reshaped the business, competing on quality control, vetted expert labelers, and enterprise-grade service. Cheap, simple work drifted to AI; high-value work went to the specialists. The generic crowd platform got squeezed from both sides until there was little room left in the middle.

Worth noting: this news has barely registered in the usual online circles. There was no dramatic shutdown announcement — just a quiet decision to stop onboarding new customers. For a service that ran for two decades, the exit is strangely muted. Then again, a silent fade may be the most fitting ending ghost work could have.

The Takeaway

The retreat of Mechanical Turk is more than one product winding down. It is a signal that an era is ending — the era in which humans trained AI with their own hands, one HIT at a time. There was always someone’s labor behind the tools we use so casually, and now AI is starting to fill even that role. Which leaves an open question: in a world where AI generates its own training data, where do the people behind the screen go? And did we ever bother to learn their names.