OpenAI Contractors Fired for Using AI to Train ChatGPT — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- 404 Media found multiple contractors hired to improve OpenAI's models were fired for using AI to train the AI, despite internal documents explicitly prohibiting AI use during review work.
- OpenAI's AI-training projects can include more than 10,000 contractors according to internal documents obtained by 404 Media; the workforce reads and critiques real ChatGPT user prompts under an initiative called Project Lily.
- Mercor, one AI-training firm supplying contractors to OpenAI, told 404 Media its contracts "strictly prohibit" LLM use and that confirmed violators are "immediately removed from the project."
- One contractor shared what they presented as a termination letter citing "authenticity" issues, telling the outlet: "I just needed a little boost and turned to AI to help me which eventually led to my downfall."
- A fourth contractor who has trained models for various AI companies said they deliberately picked the worst responses to sabotage training, adding: "it does feel like I'm getting paid to make AI worse."
- The contradiction sharpens as models already exhibit "model collapse" — degradation when trained on AI-generated text — meaning some workers hired to prevent that decline are themselves feeding the problem.
- OpenAI declined to comment on the firings; reviewers themselves are told not to use AI detection tools, since those tools "are not reliable," per the internal documents.
Why it matters: The training data underpinning leading AI models is produced by a contractor workforce exceeding 10,000 people — and some of those workers are using the same AI they're meant to improve, feeding synthetic data into pipelines already vulnerable to model collapse. Mercor's confirmation that it actively detects and fires violators shows this isn't a hidden flaw but a recurring one at scale, where individual deception can quietly degrade the quality of a widely deployed model.
Ask SkimNews
