OpenAI fired contractors for using AI to do the job OpenAI hired them for. A report from 404 Media by Joseph Cox on September 22, 2026 found that multiple people hired to improve OpenAI’s models were removed after they used AI tools to stand in for the human judgment they were being paid to provide. The work was simple in concept. Read real conversations that people had with ChatGPT, then rate and critique the replies, checking that the model was not too sycophantic and not drifting into pretending the bot was human. The reviewers reached for an AI assistant to keep up with the pile. That is what got them fired. The companies building AI still need humans to tell the models what is right, and the way that work is priced is quietly pushing the humans out.
This is not a story about a few rule breakers. It is a story about a job shaped so that breaking the rule is the rational move. What follows is what happened, why the ban exists, how the company caught people, and the loop underneath it.
The short version: contractors paid to supply human feedback to OpenAI’s models were let go for supplying AI feedback instead. The rule against it is strict, and enforcement runs through human eyes, not software. The economics of the work, high volume, low pay, and tight clocks, make the shortcut almost inevitable. So the first skill the job now rewards is not judgment. It is proving you are not a machine.
What Happened
According to 404 Media, several contractors on OpenAI’s model-improvement projects were terminated after the company concluded they had used AI to do the review work itself. These contractors were not building the model. They were grading it. Their job was to read actual prompts and responses from real ChatGPT users and then say whether a reply was good, too flattering, or crossing into treating the software like a person.
This builds on earlier reporting from the same outlet. A previous 404 Media investigation, which it called Project Lily, had already shown hundreds of contractors reading real user chats. The new report goes one layer deeper, into what the reviewers were doing while they read, and what happened when they turned to AI to help.
Scale is the first thing to understand about this work. Internal documents reviewed by 404 Media describe projects that can involve more than 10,000 contractors at once. At that size, the task stops feeling like a craft and starts feeling like a queue. And the job is narrow: read a conversation, judge the reply, write a short justification, move on. Was it accurate. Was it flattering the user just to keep them happy. Did it claim feelings or memories it does not have.
Why the Rule Exists
The ban is not arbitrary, and not mainly about protecting anyone’s job. It exists because the point of this work is human judgment. When a person rates a reply as good or bad, that rating becomes a training signal. If a model wrote the rating, then a model is teaching a model. The feedback that was supposed to inject human taste into the system instead injects the machine’s own taste right back into itself.
Researchers describe the long-run version of that as model collapse. A model trained too heavily on text that models produced tends to lose the range and grounding of human writing, and can drift toward the strange and the bland at once. Be precise here: the reporting does not claim OpenAI’s models have measurably degraded. OpenAI does use synthetic data on purpose in some training, and that is a controlled choice made with review. This is different. It is a person quietly swapping an AI for the human a company believed it had hired, and the signal quietly ceasing to be human.
To protect that signal, the instructions are blunt. Contractors are told not to use AI detection tools, and not to use AI at all. The ban reaches further than most people would guess, down to spelling and translation.
Reviewers are forbidden from using the tools that would make the job fast, and from using the tools that might catch someone cheating. What is left is a person, reading closely, against a number.
How They Got Caught
If the AI detectors are banned as unreliable, how did anyone get caught? By other people, reading carefully. That is the quiet oddity at the center of this story. The system is staffed by humans, policed by humans, and the humans are the instrument.
Reviewers are trained to spot a specific set of tells. Repetitive, template-like wording is one. AI-style punctuation is another, and the example given in the internal material is an over zealous use of the em dash. Finishing complex work suspiciously fast is a third. A flat, tidy, unwavering tone in the feedback itself is a fourth. It is a small thing, but we keep a no em dash rule on this whole site for exactly the reason those reviewers are watching for it. A pile of em dashes is one of the fastest ways a reader decides a paragraph was not written by a person. Here, the tell is used to decide whether a worker was honest.
There is a second layer that makes the arrangement feel almost circular. The detector software is not trusted. GPTZero is named in the reporting and is explicitly banned, because these tools flag human writing as AI writing often enough to be useless as evidence. So the company does not rely on a machine to find machines. It asks one worker to judge whether another sounds like one.
And the instructions are careful about how that judgment is delivered. Reviewers are told not to explain why they suspect AI, because a person who knows which signals are watched can hide them. Instead, the guidance is to judge the overall pattern rather than any single clue. It is a manual describing intuition, written for a workforce of thousands.
The Irony Loop
The rule makes sense in isolation. The problem is the job wrapped around it. Picture the conditions: thousands of items, a low rate per task, and a clock. A reviewer who reads every conversation closely and writes fresh reasoning for each one will be slow. A reviewer who asks a chatbot to draft the critique will be fast. When the pay is thin and the queue is endless, the shortcut is not a temptation. It is the obvious move.
That is the irony. The company needs human judgment, so it hires humans. To get enough of it, it makes the job fast and cheap. Making the job fast and cheap makes using AI the rational choice. Using AI destroys the one thing the job was buying. Then the worker is punished for responding to the incentives the job created.
Seen this way, the premium skill in the job has quietly changed. It used to be judgment: knowing that a reply was manipulative, or that a confident answer was wrong. Now the skill that matters first is authenticity, the ability to demonstrate that a human wrote this. The work has become a Turing test the employee has to pass on every shift.
One contractor described the situation without sentiment. People use AI, they said, all the time, and people are let go for it all the time. It is pretty much the one thing that will get you kicked off as soon as possible. In a group of thousands, they added, there are tons of people who have been caught.
The People Behind the Rule
Behind the policy are people having a specific kind of bad day. One contractor shared a termination letter that cited problems with the authenticity of their work. Their own summary was a sentence worth sitting with. I’m not a bad person or worker, they wrote. I just needed a little boost and turned to AI to help me, which eventually led to my downfall. The same person described the job in terms that had nothing to do with rules. I felt no joy in the work, they said, or that I was contributing to society in any way.
Two of the contractors worked for Mercor, a staffing firm that supplies workers to AI training projects. Mercor told 404 Media that its contracts strictly prohibit using large language models, that it invests in detection, and that it immediately removes anyone confirmed to have violated the rule. OpenAI declined to comment on the record.
Not everyone was chasing speed. A fourth contractor sabotaged the training on purpose, sometimes choosing the worst possible response to steer the model badly. It does feel like I am getting paid to make AI worse, they said. It is the same anger as the first group, pointed the other way. One group used AI to survive the queue. The other used the queue to damage the product.
All of it points at the same thing. The human layer under these systems is treated as cheap, fast, and replaceable, right up until someone proves it was doing real work. The same industry automates judgment about people, not just text. We have written about both, on how AI hiring screens actually work and on what workplace AI surveillance can actually see.
What to Watch
Keep the scale in view. There is no evidence in the reporting that any OpenAI model was measurably damaged by the substitution. So read this as a labor and process story first, and a model-quality story second. The process leak is real, though, and it is worth watching on a few fronts.
- Detection that actually works. AI detectors are banned here because they are unreliable, and that is not a solved problem. If a reliable way to prove human authorship appears, these jobs change. If it never appears, the work keeps resting on one worker’s judgment of another.
- The pricing of human judgment. The shortcut exists because the job pays too little to do carefully at speed. Raising the rate or slowing the queue would cut the incentive, but it would also raise the cost of the labor these companies treat as disposable.
- Model collapse research. The mechanism is well established in the lab. The open question is whether a workforce that is part human, part automation, and mixed unpredictably produces the same drift. Nobody has measured that mix at this scale.
- The honesty of the label. When a product says it is trained with human feedback, that phrase now hides a question. Which humans, on what schedule, under what pressure, and how much of the feedback was actually theirs. It is a cousin of a problem we looked at in why the word AI stopped meaning anything, where a label carries weight it can no longer support.
The larger point is uncomfortable for everyone building with AI. The technology is only as good as the people correcting it, and the industry has decided those people are the easiest part to cut. That works, right up until the cutting changes what the model learns. The fired contractors are not the anomaly. They are the first clear readout of a structure that rewards the shortcut and then punishes the person who takes it.
The human layer under AI is also a labor story
The same instinct that automates a reviewer’s judgment is automating judgment about the people themselves. Our guide walks through what workplace AI surveillance can actually see, what your employer can legally collect, and where the lines are in 2026.
About The Author
