{
  "id": 7048,
  "url": "https://arxiv.org/abs/2603.18280v3",
  "title": "Detection Is Cheap, Routing Is Learned: Why Refusal-Based Alignment Evaluation Fails",
  "summary": "Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: politi",
  "authors": "Gregory N. Frank",
  "category": "research",
  "topics": "regulation,safety-alignment,healthcare",
  "orgs": null,
  "regions": "china",
  "published_at": "2026-03-18T20:54:34.000Z",
  "fetched_at": "2026-07-14T16:32:54.536Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/7048",
  "original_url": "https://arxiv.org/abs/2603.18280v3",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}