{
  "slug": "human-feedback-integrity",
  "title": "The human feedback training your model may not be human",
  "summary": "Crowd workers route tasks through models, a fifth of one conference's reviews were machine generated, and MTurk is closing. Preference data needs a receipt.",
  "lede": "Reinforcement learning from human feedback rests on one assumption nobody verifies: that the human was human. Published research has found crowd workers routing text tasks through language models, a major conference found a fifth of its reviews machine generated, and the platform that created this market is shutting down over data integrity. Here is what a per judgment receipt would change, and the one thing it cannot fix.",
  "date": "2026-09-08",
  "reading_time": "18 min read",
  "category": "Research",
  "tags": [
    "RLHF",
    "data labeling",
    "proof of human work",
    "training data provenance",
    "AI evaluation",
    "preference data"
  ],
  "image": "https://cdn.twc.sh/images/igcache/Human%20Feedback%20Integrity/1200_630/blog.jpg",
  "url": "/blog/human-feedback-integrity.html",
  "wordcount": 3893,
  "related": [
    "proof-of-human-work",
    "verified-work-passport",
    "identity-failure-map"
  ],
  "schema": "Article"
}