OpenAI contractors fired for using AI to do the human AI-training work

Contractors paid to improve ChatGPT with human judgement have been dismissed for letting AI do the work instead, according to an investigation by 404 Media. The outlet spoke with three contractors working across several OpenAI projects and obtained internal documents describing both the rules governing the training workforce and how they are enforced.

OpenAI relies on a large contractor pool to refine its models, and one internal document seen by 404 Media indicates that a single project can involve more than ten thousand people. Their job, as 404 Media described in earlier reporting on the effort it called Project Lily, is to read real ChatGPT user prompts and conversations — which can include personal information — and then rate and critique the chatbot's answers, checking among other things that responses are not too sycophantic and do not present the bot as more human than it is.

The entire premise is that these workers supply something a model cannot. According to 404 Media, a notable number of them do not. One contractor said AI use happens constantly and that being caught for it is essentially the one sure way to be removed from a project, adding that within a group of thousands, many people have been caught.

What the internal documents say

A document covering contractors whose job is to review other contractors — including catching them using AI — instructs: "Do not use AI detection tools, or AI yourself." It dismisses GPTZero and similar detectors as unreliable, and counts Grammarly and AI translation tools as prohibited AI use when writing reviews, feedback or comments.

The same guidance tells reviewers not to explain what made them suspicious, on the reasoning that evaluators can hide their habits more easily if they know what is being checked, and to judge the overall pattern of someone's work rather than a single clue.

How reviewers spot AI use

Reviewers are told to watch for hallmarks of machine-written text, 404 Media reports: repetitive vocabulary, AI-typical punctuation — including overuse of the em dash — and tasks finished suspiciously fast. In Slack channels where workers ask each other for advice, one contractor said people regularly post a sample and ask whether it looks AI-generated. "Usually the answer is yes," the person said.

One contractor admitted to 404 Media that they had used AI while helping train OpenAI's models and shared a termination letter citing problems with the "authenticity" of their work. They described turning to AI because they needed a boost, and said the work gave them no sense of contributing to anything.

Mercor confirms removals, OpenAI declines to comment

Two of the three contractors work through Mercor, a company that hires people to review ChatGPT-related material. In a statement, a Mercor spokesperson said its experts are hired for their expertise and judgement, that contracts strictly prohibit using large language models to complete projects, and that anyone confirmed to have done so is removed immediately. The company added that it invests heavily in tools and systems to detect misuse.

OpenAI declined to comment on contractors being fired for using AI.

Sabotage adds another wrinkle

404 Media also spoke with a fourth contractor who has trained models for various AI companies and said they sometimes deliberately pick the worst responses in an effort to sabotage a model's training. Unsure how much one rater matters among hundreds of others, the person said it still feels "like I'm getting paid to make AI worse."

Why it matters

Human feedback is the ingredient that is supposed to make chatbots helpful and safe, so its value depends entirely on being genuinely human. If a slice of it is machine-generated, the training signal degrades exactly where it should be strongest — and researchers already warn about model collapse, the compounding decline that can occur when models are repeatedly trained on AI-generated text. These raters are the people hired in part to prevent that outcome. The reporting also shows how fragile enforcement is: automated detectors are considered too unreliable to use, leaving detection to manual pattern-spotting across workforces of thousands. And there is a broader irony, as 404 Media points out: OpenAI is spending heavily to persuade businesses to put AI in every workflow, while banning it outright in its own training pipeline, where the human touch is the entire point.