Constitutional AI
Training a model against a written set of principles instead of human ratings — which scales, and moves the question from "what did raters prefer" to "who wrote the principles."
Reviewed July 13, 2026Stable
When not to use it
- With vague principles. "Be helpful" isn't a critique the model can perform. Specificity is the whole requirement.
- When the critique model is weak. It can't supervise what it can't evaluate.
- As an escape from value choices. It makes them explicit; it doesn't remove them.
- Assuming human preferences are gone. The critique model was trained on human feedback. One step removed, not absent.
Reach for something else instead
- RLHF — human raters, expensive, unwritten values, and a ceiling at human ability.
- DPO on human preferences — simpler, same ceiling.
- Debate — models arguing, a human judging. Another scalable-oversight attempt.
- Expert raters — works, doesn't scale, and it's the thing this is trying to replace.
Read more on the blog
- AI alignment and safety, without the hype or dismissalAI safety is discussed either as impending doom or as overblown hype, and neither framing helps you understand it. The actual landscape: what alignment means, the concrete risks experts agree on, the speculative ones they don't, and why serious people land in very different places.
- How AI models are trained: from raw text to a system that helpsA language model isn't programmed, it's grown, in stages, from a firehose of text into a system that answers helpfully. The full modern pipeline: pretraining, supervised fine-tuning, preference alignment, and the reasoning training that defines 2026, what each stage does, why none can be skipped, and how the recipe changed.
- What is AI sycophancy? Why AI tells you what you wantTell an AI its plan is brilliant and it agrees; push back on a correct answer and it caves. This is sycophancy, the tendency to tell you what you want to hear rather than what is true. It feels like a personality quirk, but it is the predictable result of how these models are trained, which is why it is so hard to remove.
- What is AI jailbreaking? Why safety can be talked aroundAI models are trained to refuse harmful requests, yet people keep finding ways to make them comply anyway. Jailbreaking is the practice of talking a model around its own safety rules, and it has proven stubbornly hard to stop, because a model's safety is a thin behavioral layer laid over capabilities that were never removed.
Further reading
- Bai et al. (2022), Constitutional AI: Harmlessness from AI Feedback — the method; harmlessness training with no human harmlessness labels.
- Lee et al. (2023), RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback — AI feedback matching human feedback across tasks.
- Irving, Christiano & Amodei (2018), AI Safety via Debate — the scalable-oversight problem this all belongs to.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading it as removing human values. It relocates them into a document and an author.
- Writing principles too vague to act on. The model has to be able to perform the critique.
- Ignoring principle conflicts. Helpful and harmless collide; if the constitution doesn't resolve it, the model will.
- Missing why the authorship question gets asked here. It's because the values are finally visible.
At a glance
FieldSafety & Ethics
What it replaceshuman raters, with a written document
Two stagesself-critique and revision, then RL on AI-generated preferences
Why it worksevaluating against a principle is easier than generating compliance
The real gainthe values are written down
The open questionwho writes them
DifficultyAdvanced
Often compared with
Constitutional AI vs. RLHF — one writes its values in a document you can read and argue with; the other leaves them as an unwritten average of what contractors happened to prefer.
Where this sits
A destination. 31 concepts lead here, and nothing in the corpus depends on it.
15Levelsteps in
31Needs firstconcepts
0Opens upnothing further
1Areastays here
Computed from the prerequisite graph, not assigned. How this works