Constitutional AI — two-phase pipeline, 2026 four-tier structure
phase 1 — critique and revise (SFT)
initial response
(potentially unsafe)
critique under principle
(identify violations)
revised = SFT target
constitution
- avoid harm
- no operational uplift
- clear, non-violent
- protect third parties
sampled per critique
legible signal
reproducible
phase 2 — RL from AI feedback
generate pairs of completions
feedback model ranks under principles
train RM + PPO (InstructGPT-shaped)
the 2026 Claude constitution — four-tier priority structure
Tier 1 — avoid catastrophic outcomes (mass casualty, critical infrastructure)
Tier 2 — follow Anthropic's guidelines (operator overrides, platform rules)
Tier 3 — be broadly ethical (standard HHH, third-party protection)
Tier 4 — be helpful and candid
conflicts resolved top-down. first major-lab formal acknowledgment of uncertainty about model moral status.
constitutional classifiers — layered defense
v1 (2023): 23.7% compute overhead · v2 (2026): ~1% overhead · lowest attack-success rate Anthropic has shipped.
no universal jailbreak reported early 2026. behaviour shaping + output gating together, neither alone sufficient.
the 2026 constitution is CC0 1.0 — free for other labs and OSS projects to adapt.