OpenAI is paying thousands of pros to teach ChatGPT their jobs

If you still picture AI training as people clicking boxes to identify traffic lights, OpenAI has moved miles past that.

Business Insider revealed an internal OpenAI effort known as "Project Stagecraft." Working through data-labeling platform Handshake AI, OpenAI has enlisted roughly 3,000 to 4,000 freelancers, paying them at least $50 an hour to teach ChatGPT how to handle complex, occupation-specific tasks.

Teaching ChatGPT the trade

Documents reviewed by BI include a 439-row spreadsheet listing a wild range of professions. We're talking commercial airline pilots, soil and plant scientists, pharmacists, composers, emergency physicians, and even sculptors.

Handshake's project manual specifies that the focus is on "knowledge work, not manual labor." The explicit goal is to map economically relevant tasks and figure out what ChatGPT can already pull off in the real world.

Contractors are instructed to create professional personas and draft realistic digital workflows. For instance, someone playing a nurse practitioner might set up context, references, and deliverables for a 10-page medical literature review. Another prompt asks the model to act as a theatre financial manager, using an excerpt from a collective bargaining agreement to build an Excel payroll sheet for touring musicians. Contractors are told never to dumb down the requests for AI.

Every piece of work goes through two rounds of review at Handshake—including one by an industry expert—before OpenAI runs its own third review.

The irony of training your replacement

Handshake AI originally started as a job platform for young professionals before pivoting into the surging data-labeling market. On its site, top-tier specialists can fetch up to $500 an hour.

Yet the human side of the project feels uneasy. As one contractor put it to BI: "We all were aware that we were basically training AI to replace us."

There are operational cracks as well. The project spreadsheet, which contained personal email addresses of hundreds of contractors, was left accessible to anyone with the link. Last month, BI also reported that several contractors had thousands of dollars in earned wages withheld after Handshake accused them of rule violations with zero right to appeal. Both OpenAI and Handshake declined to comment.

Why it matters

For anyone using ChatGPT for real work, this points to a massive impending jump in specialized competence. Generic chatbots love spitting out surface-level textbook answers that crumble the second you drop them into real industry software. By feeding the model actual digital workflows from thousands of working pros, OpenAI is teaching ChatGPT how knowledge work genuinely gets done. If your day job involves domain-specific deliverables, it's worth testing how much closer ChatGPT is getting to your day-to-day work.

For the AI landscape, the message is blunt: the era of cheap, low-wage data annotation is giving way to high-dollar harvesting of human domain expertise. Pre-training on public internet text has hit diminishing returns; the new competitive moat belongs to whoever can catalog real-world occupational workflows first. Rivals like Anthropic and Google will have no choice but to ramp up similar professional networks.

We think Stagecraft is tactically brilliant, but the friction points are glaring. Relying on contractor platforms with opaque payment enforcement and sloppy data security will burn goodwill fast. More broadly, there is a stark reality in watching thousands of white-collar specialists sell their workflow blueprints for $50 an hour. It accelerates the very economic disruption OpenAI is supposedly preparing for.