Multiple AI training contractors working for OpenAI have been terminated after using generative AI tools to execute their human feedback and evaluation duties. The contractors were hired to rate, critique, and improve ChatGPT’s outputs to prevent issues like model collapse, sycophancy, and anthropomorphization. Internal documents reveal that OpenAI strictly prohibits review personnel from utilizing AI detection software, Grammarly, or AI translation services in their workflow.

According to internal documents and worker testimonials, reviewers are instructed to flag subtle indicators of AI usage, including repetitive phrasing, specific punctuation patterns, and unnaturally fast completion times. While thousands of contractors work across these data pipeline initiatives, OpenAI relies on human oversight to keep training sets clean, making automated output in reviews grounds for immediate dismissal.

This trend underscores the systemic challenge of maintaining high-quality human data pipelines for frontier LLMs. As model collapse remains a technical risk when models train on synthetic text, reliance on outsourced labor creates operational friction when workers use AI to accelerate their output.

Why it matters

  • Data hygiene risks: Synthetic text generated by contractors can contaminate human feedback loops, accelerating model collapse in production LLMs.

  • Operational oversight: AI startups must implement strict audit processes and clear metrics to ensure outsourced human feedback remains truly human-generated.

  • Labor management: Over-reliance on low-cost contractor labor creates misaligned incentives where workers use AI to meet high quota demands.

Source: 404media.co