Freelancers working on an OpenAI project are creating task data by job category, from agriculture and livestock to music composition and civil aviation piloting, and feeding it into ChatGPT training, it has emerged.

Business Insider reported on April 1 local time that the work is called "Project Stagecraft" inside data labeling startup Handshake AI. The project is under way using about 3,000 to 4,000 freelancers, it said.

Work guidelines obtained by the outlet say contractors should set a "job persona" based on industry experience and then write prompts as if asking a colleague to do real work. They are instructed to spell out the persona's context and goals, references and deliverables, and ensure the steps leading to the deliverable can be performed only in digital form. One example involves setting a nurse specialist persona and specifying a 10-page review of medical literature on a particular disease as the deliverable.

A project training guide sets its focus on "knowledge work, not physical labor". It states the collected data are "used to map economically relevant tasks and evaluate the model's capabilities". It also sets a principle of not adjusting the level of detail based on an assumption of what "AI will need", and instead reflecting actual workflows and professional expectations as they are.

One contractor said freelancers receive at least $50 per hour, or about 76,000 won. Handshake's website says expert contract work can pay up to $500 per hour, or about 759,000 won, based on postings not directly linked to OpenAI.

Review is conducted in three stages. Handshake AI reviews contractors' results twice, including once in which an industry expert checks job-specific details. OpenAI is reported to handle the third review.

The work shows ChatGPT training has moved beyond accumulating general knowledge to a stage that reflects workplace methods and professional job language in real industrial settings. An interpretation has emerged that the scope of generative AI use is expanding further, as task data in subdivided fields such as agriculture and livestock, healthcare and aviation are being structured and fed into the system.