OpenAI contractors fired for using AI to generate the feedback meant to train OpenAI's models
2026-09-22
404 Media reported Sept 22, 2026 that multiple contractors hired to help train and review OpenAI's models — including some working through Mercor — have been fired for using AI tools themselves to generate their review feedback, based on internal documents and interviews with three contractors. Internal guidance seen by 404 Media instructs reviewers not to use AI detection tools, Grammarly, or AI translation, and not to tell flagged evaluators why they're suspected so they can't learn to evade detection; one project alone involves more than 10,000 contractors. One contractor said 'in a group of thousands there are tons that have been caught,' and one fired contractor shared a termination letter citing issues with the 'authenticity' of their work.
Significance 3: A well-sourced investigative finding (internal documents plus multiple named contractor accounts) about integrity problems inside OpenAI's human-feedback pipeline — a genuine operational-lane development even though it doesn't move a capability score directly.
Implications · machine-drafted, not owner judgment
If a meaningful share of the human reviewers meant to keep AI-generated drift out of OpenAI's training data are themselves using AI for that work, the human-in-the-loop safeguard against model collapse and quality drift is weaker than OpenAI's public framing implies — a stated-vs-revealed gap in the reliability of the review pipeline itself. For the owner's book argument about what AI structurally cannot replace (honest human judgment, accountability), this is a sharp real-world case: the layer meant to supply that human judgment is eroding under cost and scale pressure.
- OpenAI or Mercor commenting publicly on the scale of AI-assisted contractor terminations
- Evidence of measurable quality drift in ChatGPT responses traceable to this pipeline
Sources
- press People Training OpenAI's AI Fired for Using AI to Train the AI retrieved 2026-09-22