OpenAI employs a substantial number of contractors to analyze real user interactions with ChatGPT, aiming to refine the chatbot's performance. These contractors play a crucial role in humanizing OpenAI’s models. However, as uncovered by 404 Media, some of these workers have been dismissed for employing AI to assist in their tasks. This situation is particularly ironic given that OpenAI advocates for the integration of AI into workplaces.
Some of the AI models are already showing early signs of “model collapse,” a phenomenon where models trained on outputs generated by other AIs experience reduced performance. Alarmingly, individuals tasked with mitigating this issue have resorted to AI-generated content to aid in training.
One contractor stated that the use of AI is prevalent among their peers and that being caught using it often leads to immediate termination. “In a group of thousands, many have faced consequences for this,” they explained.
In a recent report, 404 Media introduced Project Lily, where hundreds of contractors scrutinize real user prompts and interactions with ChatGPT, which may contain sensitive data. These individuals evaluate and critique ChatGPT's outputs to ensure that they strike the right tone and avoid excessive flattery or anthropomorphism. This reporting was based on internal documents and insights from a contractor.
Moreover, 404 Media has accessed additional internal documents and interviewed three contractors engaged in various OpenAI projects, with estimates suggesting over ten thousand contractors are involved overall.
According to one internal document, the use of AI is strictly prohibited for contractors. It states, “Do not utilize AI detection tools or any AI in your work.” In fact, they are specifically instructed not to employ tools like GPTZero, labeling them unreliable. They should focus on the broader patterns of work instead of individual clues indicating AI use.
All interviewed contractors confirmed that they have been explicitly told not to use AI in their evaluations. Two sources indicated that individuals had been let go for violating this rule. The anonymity of these contractors was preserved as they were not authorized to communicate with the media.
Contractors tasked with assessing the work of others are on high alert for indicators of AI assistance, such as repetitive phrasing, overly formal punctuation, and unusually swift task completion. Within internal Slack channels, it's common to see requests for advice on whether a particular output is AI-generated, with responses often confirming it is.
One contractor shared their experience of using AI during the training of OpenAI’s models, revealing their termination letter cited concerns about the “authenticity” of their contributions. They expressed remorse, stating, “I didn’t feel fulfilled or like I was positively contributing to society.”
Two contractors interviewed by 404 Media were associated with Mercor, a company responsible for hiring evaluators for ChatGPT-related content. A representative from Mercor disclosed, “Our specialists are selected for their judgment, which is vital for AI development. Our contracts explicitly forbid the use of language models for project tasks, and we rigorously enforce these standards. When we identify a contractor using AI, we remove them instantly.”
Additionally, a fourth contractor who has contributed to model training for various AI firms revealed they intentionally selected low-quality responses, aiming to hinder the training process. “Initially, I felt guilty about this work,” they admitted. “Now, I either disregard the outcomes altogether or deliberately pick the worst results. I doubt it makes a significant impact given the number of reviewers, but it feels like I’m being compensated to degrade the AI.”
OpenAI did not provide any comments regarding the firing of contractors for using AI.
