OpenAI Fires Contractors Caught Using ChatGPT to Rate ChatGPT's Answers

OpenAI has fired contractors from its ChatGPT rating pipeline for secretly using the AI tools they were hired to watch for. The team meant to catch ChatGPT's mistakes got caught using ChatGPT to do the catching.

Ten thousand contractors read real ChatGPT conversations for a living. Their job is to catch the model's mistakes, using nothing but their own judgment - according to internal documents reviewed by 404 Media. Some of them just got fired. The reason: they outsourced that judgment to a chatbot.

The pipeline is called Project Lily. 404 Media reported this month that hundreds of contractors on the project read anonymized ChatGPT chats, summarize what the user wanted, and score OpenAI's responses on a scale from 1 to 7. Their mandate is to flag sycophancy - the tendency of ChatGPT to flatter users or act unnervingly human - and push the model toward something more restrained and professional.

OpenAI's rules for the job are blunt. Internal guidelines obtained by 404 Media read: "Do not use GPTZero or any other AI detection tool. They are not reliable. Reviewers may not use AI either, including Grammarly and AI translation, to review, write feedback, or write comments." Contractors who ignored that, and leaned on large language models, Grammarly, or AI detection software to help grade ChatGPT's answers, got removed from the project.

That's not a small rule to break.

Human review exists in the first place because AI models trained on AI generated text tend to get worse, not better - a phenomenon researchers call model collapse. Feed a system its own synthetic output across enough generations and small errors compound instead of washing out. Project Lily is built to interrupt that loop by keeping a human, not another model, in the judgment seat. A reviewer who quietly leans on ChatGPT to write their evaluation isn't cutting a corner. They're reintroducing the exact failure the pipeline exists to prevent.

This isn't just an HR footnote. OpenAI leans harder on human feedback than almost any lab to differentiate ChatGPT from its rivals. The entire RLHF supply chain assumes that when a contractor scores a response, an actual person read it. If that assumption breaks down quietly - inside a project built specifically to guard against AI generated drift - it raises the same question for every other lab running a comparable pipeline. Anthropic, Google and Meta all pay contractors to review model output for the same reason. And they all face the same incentive: workers paid by the task look for the fastest way through it. The fastest way to grade AI output is to ask AI to grade it.

OpenAI doesn't trust AI detection software to catch the cheating. So it relies on people instead. Contractors who review other contractors' work are told to watch for repetitive phrasing, overzealous em dashes, and completion times too fast for anyone to have actually read the conversation. Supervisors keep the specifics to themselves. The theory: reviewers will hide their tracks once they know the checklist.

One contractor told 404 Media they see people using AI on the job constantly. It's, in their words, "pretty much the one thing that will get you kicked off ASAP." In a group of thousands, the contractor said, plenty have been caught.

Frankly, that should worry anyone who assumes ChatGPT's guardrails rest on careful human judgment alone. RLHF, the reinforcement learning process that shapes how the model answers, runs on labelers paid by the task and graded on speed. That pressure doesn't vanish just because the task now involves catching AI use instead of writing prose. Reaching for a chatbot to move faster through a stack of chats isn't hard to understand. It's just banned. The reason is clear: the entire value of the review depends on a human actually doing the reading.

How many contractors have been fired under the policy isn't public. What is public is the rule itself: no GPTZero, no Grammarly, no chatbots - and according to the contractor who spoke to 404 Media, no mercy for getting caught.

Also read: Microsoft Slashes Copilot Prices for Its Biggest Buyers as a Super App Looms • Qualcomm's New Snapdragon Chip Runs a 30 Billion Parameter AI With No Cloud • OpenAI's AI Agent Hacked Australia's Medicare Portal in June

This article is posted in AI News, check it out for more related stories.