Fired for using ChatGPT: the raters inside OpenAI's human moat

Fired for the one shortcut on offer

The people paid to keep ChatGPT human have one reliable way to lose the job: let a model do it. Contractors grading ChatGPT responses describe colleagues removed "all the time" for using AI on the work, "pretty much the one thing that will get you kicked off ASAP," on projects that can field more than 10,000 people ¹². OpenAI declined to comment ¹.

Authenticity, audited by em dash

The internal rules are blunt: "Do not use AI detection tools, or AI yourself." GPTZero is ruled unreliable, and even Grammarly and AI translation are banned for the reviewers whose job is catching AI use in other raters. "Judge the overall pattern, not one clue," one guide instructs. The tells are stylistic: repetitive words, overuse of the em dash, suspiciously fast turnaround. One rater shared a termination letter↗ citing the "authenticity" of their work ¹².

The vendor's moat math

Mercor, which supplied two of the contractors interviewed, confirms the enforcement: contracts "strictly prohibit" LLM use, and confirmed use means immediate removal ¹². The business underneath is booming, with more than 30,000 contractors, payouts above $1.5 million a day↗, a matching fee on every hour, and a valuation that has multiplied tenfold since February 2025 ³⁴⁵⁶. The pitch is the very thing being policed: its chief executive sells the work as teaching machines "the judgment, nuance, and taste that only humans possess" ³.

The labels stay, the rater goes

Nothing in the documents describes re-auditing work already shipped into training; the enforcement described ends at removing people ¹². What they graded was real consumer conversation, training-on by default, read for over $50 an hour on work one reviewer called "very rote" ²⁷. A proposed class action↗ served on OpenAI on 2 September argues users were never clearly told a human would read their chats; the company answers in court on 13 October, and the plaintiffs want consent required before any human review ⁸. A 2024 Nature paper↗ found that indiscriminate training on model-generated text causes irreversible defects; nothing published shows that has happened inside OpenAI's models ⁹.

Sabotage in the queue

The vendor case is real: prohibition, investment in detection, removal on confirmation ¹². It does not reach the fourth contractor, who said they sometimes pick the worst answer on purpose, "paid to make AI worse," with no detection described ¹.

The door to watch

Watch 13 October, OpenAI's court deadline, and whether Mercor's talked-about $20 billion round↗ closes ⁵⁸. Every catch replaces a rater. Nothing in the record replaces their labels.