Contractors Training OpenAI's Models Fired for Secretly Using AI to Train the AI
OpenAI urges the world to adopt AI at work, but swiftly fires contractors who use AI to train its own models.
An investigation published on September 22 by 404 Media revealed that OpenAI and its staffing contractors, such as Mercor, employ tens of thousands of workers to review real ChatGPT conversations and provide human feedback. However, multiple contractors hired to improve OpenAI's models have been fired after being caught secretly using AI to complete their evaluations. One terminated worker, whose dismissal letter cited issues with the 'authenticity' of their work, told 404 Media: '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. I felt no joy in the work or that I was contributing to society in any way.'
Internal documents obtained by 404 Media reveal that OpenAI strictly prohibits contractors from using any AI tools for reviewing, writing comments, or feedback—even banning Grammarly and AI translation. Ironically, the internal guidelines also explicitly ban the use of AI detection software: 'Do not use GPTZero or any other AI detection tool. They are not reliable. Reviewers may not use AI either... Do not tell evaluators why you suspect AI. It is easier for them to hide if they know what you look for.' Supervisors are instructed to rely on human inspection for telltale signs such as repetitive phrases, excessive em dashes, and suspiciously fast turnaround times.
Mercor, a staffing agency providing training contractors for OpenAI, confirmed in a statement that contracts strictly prohibit LLM usage and that violators are immediately removed from projects. Meanwhile, contractor frustration has also led to deliberate training sabotage. Another contractor revealed they sometimes purposefully selected the worst AI outputs because of the sheer monotony: 'I’m getting paid to make AI worse.' Researchers warn that feeding AI-generated text back into model training pipelines triggers 'model collapse,' steadily degrading model intelligence over time.
Source: 404 Media (404media.co).
BenPig thinks training AI with AI feels a lot like a snake eating its own tail.