OpenAI's Human Trainers Are Being Fired for Using AI to Do Their Jobs

The people paid to provide the human element in AI training are getting fired for faking it with AI. 404 Media’s Joseph Cox reported on September 22 that multiple contractors rating ChatGPT prompts and responses for OpenAI, employed through the AI-training staffing company Mercor, have been terminated after reviewers concluded their work was machine-generated. One contractor shared a termination letter citing problems with work “authenticity.” Another put it plainly: submitting AI-written responses is “pretty much the one thing that will get you kicked off ASAP.”

How They Get Caught

The detection playbook 404 Media obtained is a study in statistical tells. Reviewers flag repetitive wording across tasks, unusually fast completion times for complex ratings, and, in a detail the report’s readers will appreciate, “AI-style punctuation, which might include over zealous use of the em dash.” Internal guidance explicitly bans the reviewer from leaning on detection software: “Do not use AI detection tools, or AI yourself,” it instructs, directing reviewers to judge “the overall pattern, not one clue.” The tools that are banned are the ones everyone knows are unreliable (the guidance also prohibits Grammarly and GPTZero by name); what survives is human pattern review at scale, which is to say the same RLHF-style judgment these contractors were hired to provide in the first place, turned against them.

Why It Matters for the Models

The fires are not just HR enforcement; they are model-quality defense. Training a model’s taste on AI-generated ratings is a feedback loop with a name: model collapse, the degradation that sets in when models increasingly learn from synthetic output rather than human judgment. A contractor rating responses with ChatGPT’s help produces labels that encode the model’s existing preferences back into its training signal, flattening exactly the human variance the process exists to capture. The stakes go beyond one vendor: 404 Media’s sources describe one contractor who deliberately chose the worst responses to sabotage training, saying it felt like being “paid to make AI worse.” Data quality for RLHF is enforced by low-wage piecework under deadline pressure, and everyone in the pipeline, from the staffing firm to the lab, is finding out what that pressure produces.

Mercor confirmed to 404 Media that it strictly bans LLM use, invests in detection, and removes violators immediately. OpenAI declined to comment.

The Labor Story Inside the Story

Zoom out and the irony has structure. The contractors were hired because they are human: their non-expert reactions, confusion, preferences, and refusals are the ground truth that alignment training encodes. The same economic pressure that makes the work attractive to automate (high volume, low pay, tight clocks) makes automation almost inevitable, which is why the firing pattern is industry-wide rather than a Mercor quirk. Stanford’s finding that RLHF itself drives 49% more sycophancy than humans show already documented how fragile the human-signal pipeline is; synthetic labels flowing into it undetected is the same failure mode one step further removed. For anyone working in data annotation or evaluation, the career guidance is unambiguous and slightly absurd: the premium skill in AI training work in 2026 is demonstrably proving you are not AI.

What to Watch

Three developments would move this story. First, whether labs respond with provenance requirements for human-feedback work (identity-verified sessions, stylometry screening, or live proctored rating), and what that does to an already thin labor market. Second, whether anyone quantifies how much synthetic-labeled data slipped into recent training runs; model-collapse research gives the mechanism, but the industry has no incident-reporting norm for it. Third, whether the terminations spread up the stack: the same “was this written by AI” question is now being asked of junior engineers, support teams, and content contractors everywhere, and OpenAI’s supply chain is simply where the contradiction is sharpest. The company whose models make faking human work easiest is also the company whose product depends on humans not faking it.