In a quiet room somewhere in India, a physics scholar logs onto an AI chatbot and poses a deceptively complex question—not with curiosity, but with intent. He wants it to fail.
And it does.
The model fumbles when asked, “Is the luminosity of a Cepheid variable directly or inversely proportional to the time it takes the star to brighten and dim?” The kind of question that needs scientific intuition, and not just a vast memory of text.
The scholar isn’t surprised. In fact, he is pleased. His job is to challenge the chatbot. Once the model stumbles, he gets to work. He issues a set of step-by-step instructions to help it reason through similar questions more efficiently in the future.
It’s a noble endeavour. But it’s also billable.
The physicist, who declined to be named due to confidentiality agreements, is part of a growing class of AI trainers—freelancers hired by data-training companies like Turing, Mercor, and Deccan AI to find blind spots in large language models built by OpenAI, Meta, Anthropic, and Google and create the scaffolding that makes their answers more coherent.
Until recently, few outside the AI industry had heard of Mercor and its ilk. That changed in June when Meta paid over $14 billion to acquire a 49% stake in Scale AI—one of the oldest AI data-training companies.
There was some backlash, too. Within days, OpenAI and Google both cut ties with Scale AI. Competitors were quick to proclaim their own independence. Jonathan Siddharth, founder of the US-based AI-data company Turing, took to Linkedin to state, “Turing is Switzerland—the neutral, strategic partner to all frontier AI labs.” The company had just raised $111 million in January at a valuation north of $2 billion from Malaysia’s sovereign-wealth fund and other investors.
