Closing the loop: Humans training ChatGPT fired for using ChatGPT to do the work

The human touch, mediated by a machine

The army of external contractors who read and evaluate user interactions with ChatGPT for OpenAI has one strict rule: they must not use artificial intelligence to help them with their work. 404 Media found that scores of these "reviewers" were fired precisely because they ignored the ban and used AI to generate outputs meant to train future generations of ChatGPT.

This human factor is supposed to teach models naturalness and weed out undesirable traits. But when contractors, in a rush to meet quotas, substitute human input with another AI iteration, it leads to “model collapse”. This is a degradation of quality caused by the algorithm learning from its own synthetic data instead of real human feedback.

Hunters tracking suspicious speed

Documents obtained by 404 Media point to a sophisticated control system. OpenAI employs another layer of contractors whose job is to catch the cheaters. Paradoxically, these inspectors are also forbidden from using AI detection tools (like GPTZero) because they are considered unreliable. They are also barred from using AI-based translators or grammar checkers.

According to the instructions, the main signals of cheating include repetitive vocabulary, typical “AI punctuation” (including the overuse of the em dash), and, above all, a contractor submitting work suspiciously quickly. Inspectors are forbidden from telling the cheaters how they were caught, so they can't improve their cover-ups.

The burnout factory

One affected contractor commented that they didn't feel like a bad worker; they just needed relief from the monotonous work. Another admitted to purposefully selecting the worst possible responses to intentionally sabotage the models, feeling their work had no positive societal benefit.

Mercor, a company that hires contractors for OpenAI, confirmed that the rules strictly prohibit the use of large language models and that it invests heavily in detection tools. As soon as AI use is confirmed, the worker is immediately removed from the project.

When the product contradicts the process

The most interesting part of the whole incident is the paradox of the company that caused it. OpenAI builds its business on convincing the world to use AI at work to speed up and automate boring tasks. But when the people training that very technology do it, they get fired.

It exposes the vulnerability of the current AI ecosystem, which is fundamentally dependent on an endless supply of cheap, “authentic” human labor. As models demand ever more data, the pressure on processing speed grows. People naturally turn to the exact tools they are supposed to be improving to help them cope.

Lilith's verdict

The irony of the entire AI business in a single case. They promise companies that AI will free employees from tedious work, but when the workers in their own data factory do the same, they get the boot.