{
  "id": "deliberative-alignment",
  "code": "PTL-0167",
  "term": "Deliberative Alignment",
  "aliases": [],
  "category": "security",
  "definition": "Deliberative alignment is a safety-training method in which a reasoning model is taught the text of a safety specification and trained to recall and reason about the relevant policies in its chain of thought before answering.",
  "description": "Instead of learning safe behavior only from labeled examples, the model learns to consult explicit rules at inference time. Its authors reported better resistance to jailbreaks together with fewer unnecessary refusals. It is related to constitutional AI, which also trains from written principles.",
  "example": null,
  "broader": [],
  "narrower": [],
  "related": [
    "constitutional-ai",
    "reasoning-model",
    "over-refusal",
    "jailbreak"
  ],
  "introduced": 2024,
  "added": "2026-10-10",
  "sources": [
    {
      "title": "Deliberative Alignment: Reasoning Enables Safer Language Models",
      "authors": "Guan et al.",
      "year": 2024,
      "url": "https://arxiv.org/abs/2412.16339"
    }
  ],
  "url": "https://protologue.com/t/deliberative-alignment/",
  "citation": "Protologue. (2026). Deliberative Alignment. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0167). https://protologue.com/t/deliberative-alignment/"
}