Contractors Training OpenAI's Models Fired for Using AI to Train the AI

404 Media reports that several contractors hired to improve OpenAI's models have been fired for using AI in the work they were hired to do. The findings come from internal documents and conversations with contractors working across several OpenAI projects.

What the internal documents say

One document describing the job of contractors who review other contractors' output states plainly: “Do not use AI detection tools, or AI yourself.” It also rules out GPTZero and other detectors (“They are not reliable”) and bans Grammarly and AI translation for reviewing, writing feedback or comments. Reviewers are told not to explain their suspicions, because “it is easier for them to hide if they know what you look for. Judge the overall pattern, not one clue.”

According to one document, the projects involved can include more than ten thousand contractors. 404 Media spoke to three contractors; all said reviewers are told not to use AI, and two said people have been fired or offboarded for it.

The signs reviewers look for

Reviewers are trained to spot repetitive wording, AI-style punctuation, including heavy use of the em dash, and work finished unusually quickly. In internal Slack channels where contractors ask each other for advice, a common question is “Is this AI?” One contractor said the answer is “usually yes.”

Another contractor who used AI shared what they presented as a termination letter citing problems with the “authenticity” of their work. “I just needed a little boost and turned to AI to help me which eventually led to my downfall,” they said, adding they felt no joy in the job. One contractor described AI use as “pretty much the one thing that will get you kicked off ASAP,” saying that “in a group of thousands there are tons that have been caught.”

Model collapse and the company response

The issue matters because models further trained on AI-generated text can degrade, a phenomenon known as model collapse. Last week 404 Media reported on Project Lily, in which hundreds of contractors read real ChatGPT users' prompts and conversations, sometimes containing personal information, and rate the answers, including checking that they are not too sycophantic.

Two of the contractors worked for Mercor, an AI training company. A spokesperson said contracts strictly prohibit using LLMs to complete projects and that confirmed misuse leads to immediate removal from a project. OpenAI declined to comment. A fourth contractor said they sometimes deliberately picked the worst responses: “it does feel like I'm getting paid to make AI worse.”