AI doesn't think for itself: 700,000 people train it by hand

AI doesn't think for itself: 700,000 people train it by hand

How much intelligence is really inside artificial intelligence? The figure circulating in the debate is uncomfortable: around 700,000 people —mostly in India and Africa— sustain much of the work that the corporate narrative attributes to the machine. They are not programming the future. They label, filter and correct data so the model doesn't spout nonsense. Up close, the technological miracle is a high-tech call center.

The mechanism has a technical name: RLHF (reinforcement learning from human feedback). The model is first trained on vast volumes of internet text, with no human involvement. Then comes the dirty part: people who compare answers, flag which one is toxic and which is acceptable, and polish the junk until it shines. In Nairobi or Lagos, that work pays —according to the figures being discussed— less than two dollars an hour, with no psychological support and full exposure to violent or sensual content for entire shifts.

The trick that confuses training with execution

The most widespread misconception is believing someone is typing the answer on the other side. That's not how it works, and the nuance matters. The human work is slow, fragmented and batch-based: labelers are not handed a complete text to read, but short snippets, pairs of answers or classification tasks. Once the model has been tamed with that massive feedback, it generates text autonomously, recombining patterns at processor estimulante ilegal.

Hence the paradox running through the whole debate: execution is instant and cheap; training is expensive, slow and dirty. And that hidden cost —the army of precarious workers who make it possible for the chatbot not to be an internet troll with a PhD— is exactly what the singularity narrative prefers to leave out of the investor brochure.

Is this just the usual offshoring with a sci-fi story attached?

Here the analysis splits. Some argue there is no surprise at all: it's the classic pattern of globalization, the same logic as click farms and mass outsourcing to countries with cheap labor. A complex technology is invented, total automation is sold, and underneath it all, an army of poorly paid workers is needed to make the invention work. Nothing new under the sun, just better marketing.

Against that weighs the argument that the final model is genuinely useful and autonomous, and that confusing the cleanup phase with the finished product means not understanding how this works. From that perspective, AI doesn't eliminate human creativity: it reinforces it after an initial scare. The problem wouldn't be the technology, but the story wrapped around it.

The work AI creates (and who gets it)

The notion that AI "creates jobs, it doesn't destroy them" has more legs than it seems. True: it does. The question is for whom and under what conditions. Offshoring 2.0 moves the jobs not to the professionals who antiestéticared losing them, but to a contingent of workers in low-cost countries with no contracts. Technically, it is job creation. As a growth model to present in Davos, it creaks a bit.

The issue also connects with a practical question hovering over any tech hiring decision: better a local developer or an offshore one? The sensible answer depends on the project. For a cheap prototype and a quick milestone, low cost wins by a landslide. For the core of a business that must live for years, fluid communication, knowledge of the local legal framework and continuity matter more than the initial savings.

---

In the end, the big lie is not that AI is useless. It's that its miracle has 700,000 pairs of hands behind it and a brochure that prefers to call it magic.