{
  "id": 7662,
  "url": "https://arxiv.org/abs/2603.04904v1",
  "title": "Alignment Backfire: Language-Dependent Reversal of Safety Interventions Across 16 Languages in LLM Multi-Agent Systems",
  "summary": "In perpetrator treatment, a recurring observation is the dissociation between insight and action: offenders articulate remorse yet behavioral change does not follow. We report four preregistered studies (1,584 multi-agent simulations across 16 languages and three model families) demonstrating that alignment interventions in large language models produce a structurally analogous phenomenon: surface safety that masks or generates collective pathology and internal dissociation. In Study 1 (N = 150)",
  "authors": "Hiroki Fukui",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-03-05T07:46:59.000Z",
  "fetched_at": "2026-07-14T16:33:21.052Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/7662",
  "original_url": "https://arxiv.org/abs/2603.04904v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}