Anthropic introduces Constitutional AI (RLAIF)
Confirmed
Importance: major (4 of 5)The takeaway
Anthropic’s Constitutional AI trained a harmless-but-helpful assistant using AI feedback guided by a written set of principles (a ‘constitution’) instead of human harm labels.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 29 September 2026
Your AI and this story
- GPT-6 AstraIn its training data
- Claude Opus 5.5In its training data
- Gemini 4 HydrogenIn its training data
- Grok 4.7In its training data
It happened before the cutoffs of all four assistants, so it can be in their training data.
Key facts
- arXiv 2212.08073 ‘Constitutional AI: Harmlessness from AI Feedback’ (December 2022)
- Two phases: supervised self-critique and revision, then RL from AI feedback (RLAIF)
- Used in training Anthropic’s Claude models
- Anthropic published Claude’s constitution in May 2023
What happened
The model critiqued and revised its own outputs according to principles, and a preference model trained on AI judgments then guided RL.
Why it matters
Showed alignment could scale with AI supervision, making values explicit and auditable; RLAIF is now widespread.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- Constitutional AI: Harmlessness from AI Feedback (Anthropic)anthropic.com, official
- Constitutional AI (arXiv)arxiv.org, paper
Changes
- Filed