I was on a call when a contractor quietly admitted he stopped doing the work OpenAI paid him to do. He said it was faster to paste ChatGPT transcripts into another bot and write the notes from that. You could feel the moment he knew he’d crossed a line.
One reviewer I spoke to pasted whole transcripts into a different chatbot.
I say this as someone who reads internal memos and talks to the people who live in the machine’s margins: the task is dull, repetitive, and ripe for shortcuts. You’re handed user prompts and ChatGPT replies, and your job is to judge truthfulness, tone, and safety. For a contractor paid hourly, the temptation to outsource that judgment to another AI is obvious.
OpenAI employs thousands of these contractors to read and rate conversations — sometimes private, sometimes messy — and then feed those ratings back into training pipelines. The contractors I met described a monotonous workflow that invites a simple question: if an algorithm can grade an algorithm, why shouldn’t you use it to save time?
Can AI safely evaluate other AI?
You should care about the answer. OpenAI’s stance is blunt: reviewers were explicitly told not to use AI tools such as GPTZero, Grammarly, or any other AI to evaluate, translate, or comment on logs. The reason isn’t just control; it’s technical. When models train on synthetic data produced by other models, the resulting system degrades in ways you might not see until it’s too late.
OpenAI’s internal guideline reads: “Do not use AI detection tools, or AI yourself.”
I read the memo. It was strict, to the point where one contractor told 404 Media that using AI is “pretty much the one thing that will get you kicked off ASAP.” You can feel the tension: the company wants human labels and human judgment even as they build systems designed to replace human work.
Sam Altman has said publicly that AI will take jobs and that part of his role is to help destroy them. But here’s the catch: when people hand those jobs to AI quietly, the lab that built the AI faces a new kind of risk. The company forbids AI reviewers because it sees a slow corruption risk to its training signal — human judgment contaminated by synthetic answers. It becomes a mirror that feeds itself.
Why would OpenAI fire contractors for using AI?
Think about incentives. If contractors use AI to grade AI, the labels you collect are no longer grounded in human judgment. That undermines model quality, safety testing, and the trustworthiness of any moderation regime. OpenAI appears to be policing that boundary hard because the alternative is training data that drifts away from human norms.
Researchers have warned that training on model-generated content harms models.
I opened a Nature paper and the finding was stark: indiscriminate use of model-generated content in training causes defects. IBM’s work on model collapse and other published studies back the same warning — feeds of synthetic outputs slowly erode the model’s grounding in human language.
If you keep feeding a model answers invented by other models, the result is not improvement but a gradual degradation. It’s like an engine clogging on synthetic fuel: performance falls, errors compound, and the whole system grows brittle.
Does AI-to-AI training harm models?
Yes, the evidence is mounting. A 2024 Nature study showed irreversible defects when models trained on synthetic text too freely. That’s why companies ban contractors from using AI on the very artifacts they’re paid to inspect. The risk is not a single bad label — it’s a shift in the data distribution that warps future outputs.
I don’t pretend to have a simple fix. You and I can agree that the moment will force hard choices: pay humans more to read, build better tooling for reviewers, or accept a slower slide in model quality. OpenAI fired people for a reason — but if you were the contractor, would you have done anything different?