OpenAI runs a workforce of more than 10,000 contractors whose daily output is used to refine the company’s AI models. The catch is that some of those contractors were completing their work with AI, and were terminated once OpenAI caught them.
What OpenAI’s trainer program looks like
OpenAI, like Google, relies on a large pool of human raters and trainers to test and improve its systems. The current trainer count sits around 10,000, a headcount that puts the program in the same range as Google’s well-known quality rater program, which is used to test both search results and AI outputs.
Why some trainers were fired
Multiple contractors were dismissed for using AI to complete the work they were paid to do, which is, in practice, using AI to train AI. Reports from workers themselves describe a setup where the temptation to lean on a model is high, the detection surface is narrow, and the consequence of getting caught is immediate termination.
Why this is more than an HR story
The risk in this loop is real and already documented in the research literature. AI models retrained on text produced by other AI models can show signs of model collapse, a degradation in output quality that compounds over generations. If the people hired to add a human layer end up contributing AI-generated text instead, that human layer effectively disappears, and the training set starts to look like the model’s own output. The irony is the part that travels: the people whose job is to make AI better were, in some cases, replacing themselves with AI.
What this means for AI training going forward
Programs at this scale depend on trust at the individual task level, and OpenAI’s response shows where the line is being drawn. Contractors are expected to do the work themselves. Using AI to do that work is treated as a fireable offense, not a productivity tip. The size of the program, around 10,000 trainers, is a reminder that even the most automated labs still build their models on human labor, and that human labor is now the layer being policed.
How to spot this risk in any AI training pipeline
- Audit the task, not just the output. A clean result can still come from a contractor who cut corners, so look at how the work was produced, not only what was produced.
- Require provenance. Ask for raw working notes, prompts, or session logs so the human origin of each rating or label can be checked.
- Sample at random, not just on flagged work. Detection usually catches the obvious cases; random sampling catches the rest.
- Treat AI-in-the-loop as a separate risk. If your raters have access to a model, write a clear policy on when it can and cannot be used, and enforce it.
FAQ
How many AI trainers does OpenAI have?
OpenAI’s trainer program is reported at roughly 10,000 contractors, a headcount in the same range as Google’s quality rater program.
Why were OpenAI AI trainers fired?
Multiple contractors were fired for using AI to do the training work itself, essentially using AI to train the AI they were hired to help improve.
What is model collapse?
Model collapse is the documented degradation that happens when AI models are trained on text generated by other AI models, with quality dropping further over each generation.
Related coverage
This article summarizes reporting from seroundtable.com. See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.