Contractors hired to grade and improve OpenAI's models are barred from using AI on the job, and several have been fired for submitting AI-written answers.
OpenAI employs tens of thousands of contractors to read real prompts users send to ChatGPT and then rate and critique the replies. In other words, it pays people to be the human half.
Their job is to stop AI from teaching itself
Here is the background. When a model keeps training on text generated by AI, it gets worse and worse. The phenomenon is called model collapse. These workers exist partly to block that from happening.
Yet internal documents say contractors must not use AI tools at all, not even Grammarly or translation software, and must not run GPTZero or any other AI detector. Reviewers who check other reviewers are specifically tasked with catching AI use.
The tell-tale signs include repetitive wording, AI-flavoured punctuation (an over-eager em dash, for instance), and work finished suspiciously fast. In Slack channels people often paste a passage and ask, is this AI? The answer is usually yes.
Suspicion, termination letters and quiet sabotage
One contractor shared what they presented as their termination letter. The employer cited problems with the authenticity of their work. They said they were not a bad person or a bad worker, just someone who needed a little boost and turned to AI, which led to their downfall. Another contractor said they deliberately submitted the worst possible answers, and felt fine being paid to make AI worse.
OpenAI declined to comment on the firings.
Why it matters
OpenAI tells the world to hand its work to AI while firing the people who do exactly that, and the double standard exposes the weakest joint in AI training: it still runs on a large, low-paid workforce pretending to be unnecessary.
Curated from high-quality sources, with concise summaries and key takeaways.
