Constitutional AI

Alignment approach by Anthropic that explicitly trains self-observation; predicts highest baseline and lowest prompt lift.

Neighborhood — ranked by edge-count

paper

method

few-shot prompting
implements
Providing k labeled examples in the prompt to steer model behavior.
Reinforcement Learning from Human Feedback
extends
Method for fine-tuning LMs based on human preferences; mentioned as combining RL and LMs.

concept

Alignment Faking
associated_with
Core phenomenon studied: model selectively complies with training objective to prevent modification of its out-of-training preferences

institute

Anthropic
associated_with
Lab behind Claude models and Constitutional AI training approach; represents highest baseline scores and lowest prompt lift.

framework

Self-Other Overlap (SOO) Fine-Tuning
extends
The central framework proposed in this paper: aligning AI internal representations of self and others to reduce deceptive behavior
Reinforcement Learning Constitutional AI
implements
The RL stage of CAI using AI feedback to train a preference model, then RL, resulting in a policy trained by RLAIF.
Contemplative Constitutional AI
extends
Paper's proposed adaptation of Constitutional AI incorporating contemplative wisdom charter
Supervised Learning Constitutional AI
implements
The supervised learning stage of CAI where a model critiques and revises its responses, then finetunes on revisions.
Scaling Supervision
extends
Techniques that leverage AI to help humans more efficiently supervise AI.

artifact

https://github.com/anthropics/ConstitutionalHarmlessnessPaper
about
GitHub repository containing few-shot prompts, constitutional principles, and model responses.

finding

Alignment type is the only significant predictor of scores (p=0.006); architecture and parameter count do not.
cites
Kruskal-Wallis test result: Constitutional AI predicts highest baseline; roleplay/empathy training predict lowest.

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Constitutional AI: Harmlessness from AI Feedback (Bai et al. 2022b)concept0.814
Constitutional AI method whose constitutions, if changed, could trigger alignment faking
The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'.quote0.809
Defines the core concept of the paper.
Bai et al. 2022: Constitutional AI — harmlessness from AI feedbackconcept0.800
Paper on AI-feedback fine-tuning as alternative to human-feedback RLHF; cited as ref 20
Constitutional Classifiersmethod0.789
Anthropic's inference-time guardrail filtering outputs violating constitutional rules; proposed for CCAI implementation
The constitutional approach makes it easier to control AI behavior precisely and with far fewer human labels.claim0.777
Explicit principles replace large datasets of preference labels, enabling faster iteration.
Constitutional AI can train a harmless but non-evasive AI assistant without any human harmfulness labels.claim0.772
The paper's central claim, supported by findings that RL-CAI outperforms HH RLHF in harmlessness while being non-evasive.
Meta AIinstitute0.771
Affiliation of Ziyu Guo and Rain Liu.
Autonomous AI Systemsconcept0.771
Future AI that may be rational, autonomous, and possibly conscious but lack affective consciousness.