OpenAI is running a large contractor program, known internally as Project Stagecraft, in which thousands of freelancers with real-world professional experience create training material to make ChatGPT more competent at specific jobs, according to a Business Insider investigation published on March 31.
The report, based on internal documents and interviews with participants, says between 3,000 and 4,000 freelancers are involved. They come from fields as varied as agriculture, emergency medicine, commercial aviation and pharmacy, and their task is to write realistic scenarios, personas, questions and tasks that mirror what a practitioner in their field actually does.
How the work is structured
Rather than labeling data in the conventional sense, the contractors act as subject-matter experts who design the situations an AI assistant would need to handle in a given occupation and then help judge whether the model’s responses meet a professional standard. The material feeds into training and evaluation for ChatGPT across more than 400 specialized occupations, Business Insider reported.
Pay starts at about $50 an hour and rises considerably for participants with scarce expertise, with some rates reported in the hundreds of dollars an hour. The program is run through outside staffing and data vendors, a model that other AI developers also use to recruit domain experts for model training.
Training a possible replacement
Several freelancers quoted in the report expressed unease about the implications of the work. They said they understood that the scenarios they were writing could help the model perform tasks that currently require people like them, and that the money was good enough to set that concern aside. That tension, between well-paid short-term work and long-term displacement, has become a recurring theme as AI companies pay professionals to codify their knowledge.
OpenAI has publicly said that its goal is to make ChatGPT useful across the full range of economically valuable work, and the company has released evaluations measuring model performance against professionals in dozens of occupations. Project Stagecraft appears to be the supply side of that effort: a way to obtain the expert-generated examples that general web data lacks.
A growing industry
The report illustrates how AI training has become a labor market of its own. Early data-labeling work was often low-paid and outsourced, but as models have become capable of handling basic tasks, the demand has shifted toward experienced professionals who can supply specialized judgment. Whether that work ultimately complements those professionals or substitutes for them is one of the central questions about AI’s effect on employment, and the people inside Project Stagecraft are among the first to confront it directly.
