AI and the new Mechanical Turk

For about 85 years spanning the late 18th and early 19th centuries, a strange automaton called the Mechanical Turk toured across Europe and America, drawing entranced crowds wherever it went. The machine appeared to be able to play chess all on its own, though its secret was disappointingly mundane: There was a small human chess player hiding inside.

Why were so many people taken in? According to Charles Michael Carroll, author of a book on the famous automaton, this was an era in which the “mechanical magic of the industrial revolution was a topic of day-to-day marvelling”. All sorts of new machines were being invented that were capable of previously unthinkable things. So why not play chess? In other words, the disorientating pace of technological change had created an “opportunity for some mystification and much charlatanism”.

In the 21st century, the Mechanical Turk had a second act of sorts by lending its name to a “crowd labour” platform, which distributed small, low-paid tasks to a disparate crowd of freelance workers. In the 2010s, Amazon Mechanical Turk (AMT) involved humans – hidden from sight – stepping into the gaps in supposedly automated systems because nascent artificial intelligence systems were not yet capable enough or cheap enough.

In the middle of the decade, I interviewed an American warehouse worker who earned US$5 to US$7 an hour in his spare time by doing tasks like transcribing audio clips in his “man cave”. But then the story diverged from that of the original confidence trick. Because much of the work on these platforms also involved data annotation and other ways to train AI systems to become more capable on their own.

Now, the Mechanical Turk’s second act is drawing to a close. Large language models (LLMs) have become good enough to do many of those simple computer-based tasks that were once completed by human “crowd workers”. Amazon has put a note online to say AMT will be closed to new customers from July 30, 2026. “Training those AIs looks like we digged our own grave,” one worker wrote on Reddit.

So is that the end of the story? Humans pretended to be machines; then humans trained machines; and finally machines no longer needed humans? Not quite.

The demand for humans to train AI is now shifting into the physical world, which is why there are factory workers in India with cameras strapped to their heads. It is also moving into areas that require more specific professional expertise, which is why AI training companies like Mercor are offering US$80 (S$103) to US$120 an hour for “journalism evaluators” and “finance evaluators” (the latter are expected to “identify factual, aesthetic and presentation errors in documents, spreadsheets and slide decks” and “apply deep subject-matter expertise” to grade AI outputs).

Meanwhile, there are still “crowd platforms” for people who want to do lower-paid tasks, but they are increasingly focused on areas in which human-created data is still essential – for example, for tuning AI models to human preferences, or for social science research or surveys.

But now there is a new twist: Human workers are beginning to use AI tools themselves in order to complete these tasks much faster, whether that means copy-pasting text from ChatGPT into a free-text survey response box, or even using AI agents to complete their tasks.

Prolific, which promises “real human data” that is “ready in minutes”, has experimented with various ways to detect AI agents masquerading as humans, from tracking cursor movements (humans are more erratic than agents, apparently) to including “reverse shibboleth” questions that humans should not know the answer to but an AI agent probably would.

Some researchers, meanwhile, have decided to embrace the fact that AI models are becoming very good at impersonating humans. There is a new industry growing up in so-called “synthetic sampling”, in which you get LLMs to role-play as humans with various demographics and characteristics (based on real data), and then simply survey them about their opinions instead. This is admittedly much faster and cheaper than bothering to ask any real humans, but it is not hard to imagine that something important might well be lost.

It does feel as if one chapter has closed and another has opened, but I suspect this one, too, will offer opportunities for “some mystification and much charlatanism”. The era of humans pretending to be machines is over. The era of machines pretending to be humans has begun. FINANCIAL TIMES