When a farmer in Iowa asks ChatGPT how to diagnose root rot in soybeans, and a pilot in Atlanta asks it to recall the V-speeds for a Boeing 737, the same artificial intelligence delivers two highly specialized answers. But here’s what most users don’t see: behind those responses lies a sprawling, invisible workforce of thousands of freelancers, agronomists, commercial pilots, welders, and even beekeepers who have spent countless hours teaching ChatGPT what expertise actually sounds like.
Welcome to OpenAI’s “domain specialist” program, one of the most ambitious and secretive data-labeling operations in tech history.
Since late 2022, OpenAI has quietly built a global network of freelance subject-matter experts tasked with a deceptively simple job: making ChatGPT smarter by correcting its mistakes in fields no generalist could master. Unlike the early days of AI training, where low-wage workers in Kenya and Venezuela labeled images or flagged toxic comments, this new wave requires real credentials. Think flight instructors, soil scientists, and HVAC technicians.
“We’re not just checking grammar. We’re rewriting entire technical explanations so the model doesn’t hallucinate a crop rotation schedule that would ruin a harvest,” said Marcus Tull, a freelance agronomist in Nebraska who has logged over 400 hours on OpenAI’s internal training platform since March 2023. “One time, ChatGPT suggested planting corn in a field with a known pH of 5.0 without lime treatment. That’s a rookie mistake. I had to correct it line by line.”
The operation works like this: OpenAI contracts with specialized data vendors, including Invisible Technologies and CloudFactory, who recruit and vet freelancers. Workers then receive model-generated responses to highly specific prompts, such as “Explain how to recover from a wind shear encounter during final approach” or “List the steps to troubleshoot a hydraulic leak on a John Deere 8R tractor.” The freelancer scores the answer, rewrites flawed sections, and provides a model “golden response” that OpenAI then uses to fine-tune future versions of ChatGPT.
For commercial aviation training, the stakes are especially high. Elena Vasquez, a former regional airline captain now freelancing from her home in Arizona, said she’s rejected nearly 15% of ChatGPT’s aviation answers for critical inaccuracies.
“The model once confused ‘V1’ and ‘Vr’ speeds, a mistake that, if followed, could lead to a rejected takeoff after the point of no return,” Vasquez told Business Insider. “I have to assume a student pilot might read this. So I write my corrections as if someone’s life depends on it. Because in a way, it could.”
OpenAI declined to disclose the exact size of its domain-expert freelancer network. Still, internal documents seen by Business Insider suggest the company has engaged more than 2,500 specialized contractors across 45 countries. Pay ranges from $25 to $65 per hour, depending on credential level, significantly higher than the typical data-labeling rate of $3 to $15 but still well below what a professional consultant would charge.
Critics argue that OpenAI is exploiting the gap between gig work and expertise. “These are highly skilled professionals doing teaching-level work without benefits, job security, or equity,” said Dr. Mira Reddy, a labor economist at UC Berkeley. “It’s brilliant for model alignment. But ethically, it’s a gray zone.”
OpenAI spokesperson Kelsey Chen responded in a statement: “Our domain expert program is fully voluntary, transparent about rates, and designed to improve model safety. Many freelancers tell us they value the flexibility and the intellectual challenge of teaching cutting-edge AI.”
Still, the operation reveals a little-understood truth about the AI revolution: even the most advanced models remain astonishingly dependent on human judgment. Every time ChatGPT correctly navigates a niche technical question, it’s standing on the shoulders of anonymous freelancers who worked in relative obscurity.
For now, OpenAI continues to expand the program. Recent job postings target maritime navigation specialists, veterinary surgeons, and even sommeliers. Because a model that can help a cargo ship avoid a collision and recommend a Barolo for dinner isn’t magic it’s just thousands of experts, quietly fixing hallucinations, one prompt at a time.
“Most people think AI just learns from books and the open web,” said Tull, the agronomist. “But the real curriculum is written by freelancers in their living rooms. And we’re the only reason ChatGPT knows the difference between a disk harrow and a moldboard plow.”
Next time you ask it something critical, he added, “just remember: there’s probably a human behind that answer. A tired, underpaid, but deeply knowledgeable human.”
