OpenAI Contractors Hired To Humanise ChatGPT Were Caught Using AI on the Job Despite Ban and Were Fired

Multiple contractors working on OpenAI-related AI training projects were fired or removed from assignments after using AI in work where its use was explicitly prohibited, according to an investigation by 404 Media published in September.

The contractors were tasked with reviewing, rating and critiquing AI outputs, including ChatGPT responses, while rules required them to rely on their own judgement.

The findings expose an awkward contradiction in the business of training artificial intelligence.

Human workers are still needed to decide whether an AI response is useful, natural and appropriately behaved, yet some of those workers reportedly turned to AI to help complete the very tasks for which their human judgement was being sought.

404 Media spoke to three contractors working across OpenAI-related projects. Two sources told the publication that people had been fired or 'offboarded' for using AI, while one contractor said they had personally been terminated after using an AI tool.

OpenAI Contractors Were Told Not To Use AI

The prohibition was spelled out in internal guidance reviewed by 404 Media.

'Do not use AI detection tools, or AI yourself,' the instructions stated. Reviewers were specifically told not to use GPTZero or other AI detection systems and were prohibited from using AI tools, including Grammarly and AI translation, to review work, write feedback, or leave comments.

That distinction matters because being removed from a project is not necessarily the same as being dismissed from employment. Some contractors worked through third-party AI training companies, meaning the available reporting does not establish that OpenAI itself directly fired every worker involved.

One of those companies, Mercor, confirmed that its contracts prohibit the use of large language models to complete projects. A spokesperson said its experts were hired for their 'expertise and judgement' and that the company enforced the rule.

'When we confirm an expert has used AI to complete a task, we immediately remove them from the project,' the spokesperson told 404 Media.

The investigation also found that reviewers were trained to identify signs that another worker may have used AI. Contractors told 404 Media they looked for repetitive wording, AI-style punctuation and unusually fast completion times. The overuse of the em dash was reportedly among the stylistic clues reviewers were taught to notice.

The irony was difficult to miss.

Contractors also used Slack channels to discuss whether examples submitted by colleagues appeared to have been generated by AI. One contractor told 404 Media that the answer was 'usually' yes.

Another described AI use among workers as widespread, saying that 'in a group of thousands there are tons that have been caught'. That remains an individual contractor's observation rather than an independently established measure of how common the practice was across OpenAI's wider contractor workforce.

404 Media also reported that one internal document indicated the projects could involve more than 10,000 contractors. That figure refers to the potential scale of the projects and does not mean that more than 10,000 people were involved in reviewing ChatGPT responses or that they were all subject to disciplinary action.

OpenAI Contractors Help Evaluate Real ChatGPT Conversations

The latest findings follow separate 404 Media reporting about Project Lily, an OpenAI effort involving hundreds of contractors reviewing real ChatGPT prompts and conversations.

Those workers assess chatbot responses and provide feedback on issues including sycophancy and anthropomorphism.

The reporting also noted that conversations reaching reviewers can contain sensitive personal information, despite efforts to remove or reduce personal data before human review.

The use of human reviewers is important because AI-generated material creates a potential problem of its own when it is repeatedly fed back into AI systems.

Researchers have warned about a phenomenon known as 'model collapse', in which models trained indiscriminately on material generated by earlier AI systems can progressively lose information from the original data distribution. A 2024 study published in Nature found evidence of this effect under certain recursive training conditions.

That research does not show that the contractors described by 404 Media have caused ChatGPT to deteriorate, nor does it establish that OpenAI's models are undergoing model collapse. But it shows that as AI-generated material becomes increasingly common, developers have a growing interest in preserving high-quality human-generated data and human judgement during training.

One contractor who said they were terminated after using AI described the decision in bluntly personal terms.

'I'm not a bad person or worker. I just needed a little boost and turned to AI to help me which eventually led to my downfall,' they told 404 Media.

OpenAI declined to comment to the publication on claims that contractors had been fired for using AI. Mercor, meanwhile, said it uses tools and systems designed to detect misuse and removes experts when AI use is confirmed.

The reporting leaves some details unresolved, including how many contractors were actually dismissed rather than removed from individual projects.