论文卡-Constitutional AIPaper Card - Constitutional AI

研究问题（摘要开篇原话）："we would like to enlist their help to supervise other AIs"——目标是 "training a harmless AI assistant through self-improvement"，且 "without any human labels identifying harmful outputs"；人工监督只剩一份规则或原则清单（"The only human oversight is provided through a list of rules or principles"），作者据此命名 Constitutional AI。

方法证据位置（摘要自述两阶段）：监督阶段的做法是 "generate self-critiques and revisions" 再 "finetune the original model on revised responses"；强化学习阶段用模型比较两个样本、从 AI 偏好数据训出偏好模型并把它当奖励信号，摘要称之为 "RL from AI Feedback"（RLAIF）。正文里的实验数值与设置摘要一条都没给。

作者自陈的结论口径：训出的是 "a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them"；两阶段都可借思维链式推理改善人评表现与决策透明度。摘要末句 "with far fewer human labels" 说的是人工标签的数量与可控性，不是安全性保证。摘要没有独立的局限清单，局限要看正文——本次核验只覆盖 arXiv abs 页摘要。

它没有主张的：没有主张这套方法消除有害输出，也没有主张「用了 Constitutional AI 就等于安全」；原则清单由人来写，写哪些、漏哪些仍是人的判断，这一步在摘要层面没有被验证。

读法提示（虚构需求）：把它当「监督来源换成规则文本」的做法来读，而不是当成安全证书。输入是你的场景里有一批必须拒绝且要说明理由的请求 → 处理是逐条问「这条能否写成一条可执行原则，还是要靠运行时拦截」→ 结果判断是凡是只能靠运行时拦的条目，就判定这篇方法覆盖不到 → 下一步是把这些条目写进自己的复核与事件流程，别在文档里写「已用 Constitutional AI」当作结论依据。

Research question (abstract's opening): "we would like to enlist their help to supervise other AIs" - the target is "training a harmless AI assistant through self-improvement", achieved "without any human labels identifying harmful outputs"; "The only human oversight is provided through a list of rules or principles", which is why the authors name it Constitutional AI.

Where the method sits (two phases as the abstract describes them): the supervised phase is to "generate self-critiques and revisions" and then "finetune the original model on revised responses"; the RL phase compares two samples with a model, trains a preference model from AI preferences and uses it as the reward signal - the abstract calls this "RL from AI Feedback" (RLAIF). Experimental values and settings appear nowhere in the abstract.

What the abstract's own conclusion says: the result is "a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them", and both phases can use chain-of-thought style reasoning to improve human-judged performance and transparency of AI decision making. The closing phrase "with far fewer human labels" is about the volume of labels and controllability, not a safety guarantee. The abstract carries no separate limitations section; limitations live in the body, and this check covered only the arXiv abstract page.

What it does not claim: that the method removes harmful outputs, and that "using Constitutional AI" equals safety. The principle list is written by people - what it includes and what it omits is still a human judgement, unverified at the abstract level.

How to read it (fictional need): read it as "supervision source replaced by rule text", not as a certificate. Input is a scenario with requests that must be refused and explained -> processing asks per rule whether it can be written as an executable principle or needs runtime interception -> the judgement is that anything only stoppable at runtime falls outside this method's coverage -> next step: put those items into your own review and incident workflow instead of citing "we use Constitutional AI" as evidence.