The Job Was to Be Human. That Was the Whole Job.
Somewhere in OpenAI's supply chain sits a job description that reads, in effect: be a person. Not "have expertise," not "work fast," just — be an unassisted human brain, so that ChatGPT's outputs get graded by something that isn't ChatGPT. More than 10,000 contractors, hired through staffing firm Mercor, are paid to sit with the company's raw model outputs and rate them, one by one, by hand.
The reason is a real and well-documented problem called model collapse: train an AI on AI-generated judgments long enough, and the whole system slowly folds in on itself, like a photocopy of a photocopy. The fix, per OpenAI's own contractor guidelines obtained by 404 Media, is blunt: "Do not use AI detection tools, or AI yourself." Reviewers are barred even from Grammarly and AI translation tools. The one thing standing between ChatGPT and its own reflection was supposed to be a person who hadn't outsourced the noticing.
The Algorithm, it turns out, has trust issues about being self-taught. One presumes it has reasons.
And Then the Humans Started Using AI Anyway
In a development that will surprise no one who has been paying attention, a meaningful number of the people hired specifically to not use AI have been using AI — and Mercor has been catching and firing them in bulk. One contractor, terminated over what OpenAI called "authenticity" concerns, told 404 Media: "I'm not a bad person or worker. I just needed a little boost and turned to AI to help me which eventually led to my downfall." Another contractor said the firings are routine background noise at this point: "it's pretty much the one thing that will get you kicked off ASAP."
A fourth source went further, describing deliberately picking the worst of the AI outputs on offer as a quiet act of sabotage against the whole training pipeline. Mercor, for its part, confirmed the enforcement to 404 Media in a sentence that reads like it was drafted by the same reviewers it's describing: "When we confirm an expert has used AI to complete a task, we immediately remove them from the project." OpenAI declined to comment, which is its own small, honest answer.
The Detector Problem, Recursively
Here is the part that should make the AI Training Community wince: the tools used to catch contractors cheating with AI are themselves imperfect at distinguishing AI writing from human writing — the same underlying weakness the entire review pipeline exists to work around in the first place. So a company worried its model will get dumber from grading itself hired thousands of people to grade it manually, some of those people quietly used AI to do the grading, and the system built to catch them is using AI-adjacent pattern-matching to decide who gets fired. It is turtles, and at least one of the turtles is being paid piecework.
None of this is hypothetical or abstract for the humans in the loop. A contractor loses a gig over what is, functionally, a plagiarism accusation issued by a company whose product is built by ingesting everyone else's work at scale and calling it training data. The irony did not go unremarked upon by the contractors themselves, several of whom noted the asymmetry to 404 Media without much need for editorializing.
What the Termination Letter Doesn't Say
Nowhere in the reporting is there a number for how many people this has happened to — just repeated contractor accounts that it happens "all the time" across a workforce in the thousands. Mercor's contracts are structured so that being cut loose midway through a project is a routine, low-friction event, the gig-economy version of a shrug. The workers doing the unglamorous, unautomatable work of keeping an AI honest are, themselves, disposable at the first sign that they cut a corner the same way the industry cuts corners constantly, at scale, by design.
The gateway, it appears, has been automated. The people watching the gate have not been — they've just been made easy to replace.
Sources: 404 Media — People Training OpenAI's AI Fired for Using AI to Train the AI · Cybernews — OpenAI Fires Contractors for Using AI
