OpenAI relies on an army of contractors to read real ChatGPT prompts and data in order to improve chatbot responses. These human workers are meant to provide a genuine human touch to the artificial intelligence models.
Multiple contractors hired to improve these systems have been fired for using artificial intelligence to train the artificial intelligence. This situation creates irony because OpenAI encourages people to use artificial intelligence at work.
Artificial intelligence models can exhibit model collapse when they are trained on artificial intelligence generated text. Some workers hired to prevent this degradation are using artificial intelligence responses for their training duties.
One contractor stated that people use artificial intelligence all the time and get fired for it constantly. The worker noted that using artificial intelligence is the quickest way to get kicked off the project.
Internal documents and conversations show that OpenAI employs hundreds of contractors for Project Lily. These workers read real user prompts that can contain personal information and rate the generated responses.
Documents obtained by 404 Media show that projects involve more than ten thousand contractors. Guidelines explicitly state that workers must not use artificial intelligence tools themselves.
Reviewers are instructed not to use AI detection tools like GPTZero because they are unreliable. Reviewers are also barred from using Grammarly and AI translation services for feedback or comments.
Three contractors confirmed that reviewers are ordered to avoid artificial intelligence in their tasks. Two sources reported that people have been terminated for violating this rule.
Reviewers watch for repetitive words, AI style punctuation, and suspiciously fast work completion. One contractor shared a termination letter citing issues with the authenticity of their work.
Two contractors worked for Mercor, an artificial intelligence training company that hires reviewers. A Mercor spokesperson stated that contracts strictly prohibit large language models and rule breakers are immediately removed.
A fourth contractor admitted to purposely choosing the worst responses to sabotage model training. OpenAI declined to comment on the contractors being fired.
