OpenAI Contractors Fired for Using AI on ChatGPT Training Work, Internal Rules Show

SAN FRANCISCO — Contractors hired to rate ChatGPT answers and improve OpenAI's models have been removed from projects after they used artificial intelligence to do that work, according to internal documents and workers who spoke to 404 Media.

The outlet reported Tuesday that it reviewed files and interviewed three contractors across OpenAI-related assignments. One document indicated some projects can involve more than 10,000 contractors. Two of the three said people had been fired or "offboarded" for using AI. 404 Media granted them anonymity because they were not allowed to speak to the press.

One worker said colleagues use AI "all the time and people are let go for it all the time, it's pretty much the one thing that will get you kicked off ASAP," and that "in a group of thousands there are tons that have been caught." Slack channels, the person said, fill with posts asking "Is this AI?" The answer is often yes.

Reviewers who audit other contractors are told not to rely on detectors. One guideline reads: "Do not use AI detection tools, or AI yourself." Another states: "Do not use GPTZero or any other AI detection tool. They are not reliable. Reviewers may not use AI either, including Grammarly and AI translation, to review, write feedback, or write comments." Staff are also told: "Do not tell evaluators why you suspect AI. It is easier for them to hide if they know what you look for. Judge the overall pattern, not one clue." Flags include repetitive phrasing, heavy use of the em dash and unusually fast turnaround.

A former contractor shared what they described as a termination letter citing problems with the "authenticity" of their work. "I'm not a bad person or worker. I just needed a little boost and turned to AI to help me which eventually led to my downfall," the person told 404 Media. "I felt no joy in the work or that I was contributing to society in any way."

Two of the sources worked for Mercor, a firm that supplies reviewers for ChatGPT-related material. A Mercor spokesperson said: "Our experts are hired for their expertise and judgement, which is essential to the ongoing advancement of AI. Our contracts strictly prohibit the use of LLMs to complete projects and we enforce that. We invest heavily in our tools and systems to detect misuse and ensure our experts comply with project rules and contract terms. When we confirm an expert has used AI to complete a task, we immediately remove them from the project."

The work sits inside a larger human-feedback pipeline. On Sept. 14, 404 Media described Project Lily: contractors reading real ChatGPT conversations — which can include personal details — and scoring replies, including whether they are too sycophantic or treat the bot as human. Crossing Hurdles has recruited for that work; pay has been reported above $50 an hour on some assignments. Consumer ChatGPT plans have training on by default unless users opt out.

Researchers have warned that training on synthetic text can degrade models, a pattern sometimes called model collapse. A Nature paper found that indiscriminate use of model-generated content can strip information from the original data distribution. The 404 Media account does not prove collapse in OpenAI's systems. It does show the same companies that sell AI at work are paying people not to use it on the labels that train the next model.

A fourth contractor, who has trained models for multiple firms, told 404 Media they sometimes picked poor answers on purpose. "I did feel guilty about doing this kind of work at the start," they said. "I either pay zero attention to the results and choose randomly or purposely choose the [worst] output. I'm not sure how much of a difference it actually makes since there are hundreds of other people also rating prompt results, but it does feel like I'm getting paid to make AI worse."

OpenAI did not issue a detailed public reply in the 404 Media report. The contradiction is not subtle. The product is an assistant meant to draft, grade and summarize. The training job forbids Grammarly. Vendors say judgment is the point. Workers say the tasks are dull, the queues are huge and the shortcut is the same software they are paid to improve. When the shortcut is detected, the human is removed and another contractor takes the queue.