OpenAI fires contractors for using AI to label ChatGPT’s answers

Contractors OpenAI hires to label ChatGPT's answers are losing their jobs for using neural networks on that very work. On September 22, 404 Media reported that it had spoken with three such workers and obtained internal documents where the rules are spelled out in plain terms.

The work is manual: read real user requests and evaluate the model's answers, including making sure it doesn't flatter the person it's talking to or attribute human traits to itself. According to one internal document, some projects employ more than ten thousand contractors.

The documents say outright that AI is not allowed on the job. Grammarly and machine translation get their own mention. Reviewers who look for violators among their colleagues are likewise barred from AI detectors such as GPTZero, which are considered unreliable. Suspicions are to be judged on the overall picture: repeated words, "the AI-style em dash," work handed in too quickly. One of those fired was notified about problems with the "authenticity" of their work.

Two of the people the outlet spoke with worked through the company Mercor. It stated that its contracts strictly prohibit the use of language models and that confirmed violators are immediately removed from projects. One contractor said he sees this all the time: in a workforce of thousands, those caught, by his account, are not few.

The irony is that this is how AI gets trained with AI. Training models on generated text already leads to collapse: quality degrades over time. Another person who spoke with 404 Media admitted that he sometimes deliberately picked the worst answer options in order to harm the training.

OpenAI declined to comment on the firings.

Image: Original: OpenAI Vector: Zhing Za Public domain

Translation editor: Салтанов А.