OpenAI Is Firing Contractors for Using AI to Rate ChatGPT's Answers. More Than 10,000 Workers Are Bound by the Same Rule.

OpenAI has fired multiple contractors hired to rate and improve ChatGPT's responses after catching them using AI tools to do the work, according to internal documents reviewed by 404 Media. The instructions given to reviewers are explicit: "Do not use AI detection tools, or AI yourself... Reviewers may not use AI either, including Grammarly and AI translation, to review, write feedback, or write comments." Internal documents reference more than 10,000 contractors across OpenAI's various rating pipelines, employed through the labor platform Mercor, which confirmed to 404 Media that its contracts explicitly prohibit workers from using large language models to complete these projects.

One of the affected efforts, Project Lily, has hundreds of contractors read real ChatGPT user prompts and conversations and rate the model's responses — including checking for outputs that are overly sycophantic or that anthropomorphize the AI, precisely the kind of judgment call OpenAI needs a human, not another AI, to make. Reviewers are trained to spot violations through behavioral tells rather than software: repetitive word patterns, "AI-style punctuation" such as overused em dashes, and unusually fast completion times. Internal guidance instructs supervisors not to reveal these detection criteria to contractors, reasoning that "it is easier for them to hide if they know what you look for" — and separately instructs reviewers not to trust AI-detection software like GPTZero, calling such tools "not reliable," even while banning contractors from using those same tools.

The concern extends beyond fairness to workers. This exact labeling layer feeds reinforcement learning from human feedback, the technique used to shape how models like Opus 5.5 and GPT-6 Sol behave — and a well-documented failure mode known as "model collapse" describes what happens when AI systems are repeatedly trained on AI-generated content rather than genuine human judgment: performance degrades in ways that compound across generations. If contamination at the labeling layer is as routine as 404 Media's reporting suggests, it's a model-quality problem OpenAI cannot easily unwind after the fact — one that has nothing to do with the model architecture itself and everything to do with the integrity of the human feedback loop underneath it.

The rule makes practical sense on its own terms — these workers are hired specifically for human judgment, which is worth nothing if an AI is quietly doing the judging instead. But the irony sits close enough to the surface that it's worth stating plainly: the company selling AI as a tool for nearly every kind of work has concluded, in writing, that its own product cannot be trusted to do this one specific job — evaluating itself — without degrading the thing being evaluated. That's a narrower claim than a verdict on AI generally, but it's a real one, made by the company with the most at stake in getting the answer right.