What Project Lily is
404 Media reported on Tuesday that OpenAI has fired multiple contractors for using AI tools in work they were hired to do by hand. The work is Project Lily, in which hundreds of contractors read real ChatGPT users’ prompts and conversations — which can contain personal information — and then rate and critique the model’s replies.
Part of the brief is judging tone: making sure answers are not excessively flattering and do not anthropomorphise the chatbot. Internal documents seen by 404 Media put the scale of OpenAI’s contractor programmes at more than 10,000 people across multiple projects.
The rule, and how it is enforced
The instruction to reviewers is blunt. Internal documents quoted by 404 Media say: “Do not use AI detection tools, or AI yourself… Reviewers may not use AI either, including Grammarly and AI translation.”
Mercor, an AI-training company that hires contractors to review ChatGPT material, told 404 Media that its “contracts strictly prohibit the use of LLMs to complete projects and we enforce that.”
Enforcement is not a detector. Contractors described supervisors watching for repeated words, an unusual number of em dashes, and work completed unusually fast. One termination letter cited a problem with the “authenticity” of the work. A source told 404 Media that using AI “will get you kicked off ASAP,” and that “in a group of thousands there are tons that have been caught.”
OpenAI declined to comment on its contractors being fired for using AI, according to 404 Media.
Why the rule exists
The value of this labour is precisely that it is not model output. A reviewer is being paid for a judgement the model cannot make about itself: whether an answer is too sycophantic, whether it overclaims, whether it reads as a person rather than a program. Run that judgement through a chatbot and the training signal becomes the model grading its own homework, laundered through a payroll.
The irony is doing real work here, but the underlying problem is ordinary labour economics. These roles are paid by the task, remote, and measured on throughput. A tool that triples throughput is exactly what a piece-rate worker reaches for, and a rule against it is only as strong as the detection behind it — which, on this account, is a supervisor counting dashes.
What to watch
Whether any lab publishes a measurement of how much AI-assisted rating has leaked into its own preference data, and what that does to alignment work built on top of it. So far the enforcement is anecdotal and the contamination unquantified.
