OpenAI Fires Contractors for Using AI to Train Its AI Models, Exposing Industry-Wide Challenge

OpenAI Contractors Fired for Using AI in Training Work

OpenAI has removed multiple contractors from AI training projects after discovering they used AI tools to evaluate and improve ChatGPT responses, according to a report by 404 Media3

3

. The company employs thousands of OpenAI contractors who read real ChatGPT user prompts and conversations to provide human oversight in AI training1

1

. These workers are specifically hired to deliver authentic human judgment on response quality, accuracy, and appropriateness. One contractor told 404 Media that using AI is "pretty much the one thing that will get you kicked off ASAP" and that "in a group of thousands there are tons that have been caught"3

3

.

Strict AI Policy Enforcement to Preserve Human Judgment in AI Training

Internal documents obtained by 404 Media reveal explicit prohibitions against AI use. The guidelines state: "Do not use AI detection tools, or AI yourself. Do not use GPTZero or any other AI detection tool. They are not reliable. Reviewers may not use AI either, including Grammarly and AI translation, to review, write feedback, or write comments"1

1

. These restrictions extend across various projects, some involving more than ten thousand contractors3

3

. Mercor, an AI training company that hires contractors for OpenAI projects, confirmed its contracts strictly prohibit the use of large language models to complete assignments and that violators are immediately removed3

3

.

Model Collapse Risks from AI-to-AI Feedback Loops

The enforcement stems from concerns about model collapse, a phenomenon where AI models trained on synthetic data or AI-generated content progressively deteriorate2

2

. A 2024 study published in Nature found that "indiscriminate use of model-generated content in training causes irreversible defects in the resulting models"2

2

. When contractors substitute AI-generated assessments for genuine human feedback, they create AI-to-AI feedback loops that undermine the training pipeline's integrity. OpenAI's approach contrasts with its use of synthetic data in some training processes, where AI-generated information deliberately fills gaps in training datasets4

4

.

Detection Methods Rely on Pattern Recognition, Not AI Detection Tools

Reviewers tasked with monitoring other contractors are instructed to identify AI use through behavioral patterns rather than AI detection tools. Internal documents direct them to watch for repetitive wording, excessive use of em dashes, fast completion times, and overall submission patterns4

4

. The guidelines explicitly warn: "Do not tell evaluators why you suspect AI. It is easier for them to hide if they know what you look for. Judge the overall pattern, not one clue"3

3

. One contractor described how colleagues frequently post examples in Slack channels asking "Is this AI?" with the answer usually being yes3

3

.

Contractor Perspectives Reveal Motivation and Sabotage

One terminated contractor shared their experience with 404 Media, stating: "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. I felt no joy in the work or that I was contributing to society in any way"3

3

. The termination letter cited issues with the "authenticity" of their work5

5

. A fourth contractor revealed they purposefully chose worst responses to actively sabotage model training, stating they either pay "zero attention to the results and choose randomly or purposely choose the [worst] output"3

3

.

Broader Implications for Human Oversight in AI Training

This situation exposes the tension in AI development: companies promote AI adoption across workplaces while requiring human-only input for critical evaluation processes where authentic human judgment remains essential5

5

. The contractors work on reviewing ChatGPT responses as part of initiatives like Project Lily, where hundreds of workers assess whether responses are too sycophantic or inappropriately anthropomorphize ChatGPT3

3

. OpenAI's acknowledgment in previous research that models trained using human feedback can be influenced by the quality of data labelers underscores why maintaining genuine human judgment matters4

4

. The firings highlight an industry-wide challenge as AI companies increasingly depend on human feedback to improve AI models while simultaneously restricting the use of AI tools by the people providing that feedback.