Constitutional AI, first introduced by Anthropic in a 2022 research paper, was a landmark contribution to AI alignment. Rather than relying solely on human feedback to train model behavior, it used a set of principles — a "constitution" — to guide the model in critiquing and revising its own outputs. This made training more scalable and the resulting model's values more transparent and auditable. Now, Anthropic has published a substantially revised version: Constitutional AI v2, which underpins the training of Claude 4 Opus and Sonnet 4.5.
The core intuition of the original remains. The model is trained with a two-phase process: first, supervised learning on human-written examples that demonstrate helpful, harmless, and honest responses; second, reinforcement learning from AI feedback (RLAIF), in which the model itself generates critiques and revisions guided by the constitution. What changes in v2 is the depth and sophistication of every component in this pipeline.
What Changed in v2
The most significant architectural change is the introduction of hierarchical principles. In Constitutional AI v1, all principles in the constitution were treated with roughly equal weight. In v2, principles are organized into tiers: foundational safety constraints at the top (covering catastrophic harm, CBRN risks, and child safety), followed by core behavioral norms (honesty, avoiding manipulation), and finally situational guidelines that are context-sensitive. When principles conflict — as they often do in edge cases — the hierarchy resolves the tension deterministically rather than leaving it to probabilistic model behavior.
Constitutional AI v2 Improvements
- Reduction in harmful output rate vs. CAI v1−64%
- Reduction in "over-refusal" rate−41%
- Self-consistency score on values probes+38%
- Principles in the constitution78 (up from 29)
The critique-revision loop has also been significantly upgraded. In v1, the model would generate a response, critique it against the constitution, revise it, and repeat for a fixed number of iterations. In v2, the loop is adaptive: the model dynamically estimates how harmful or misaligned a candidate response is, and allocates more critique-revision cycles to higher-risk outputs. This means the model spends its "safety compute budget" where it is most needed rather than uniformly across all outputs.
"Constitutional AI v2 is not merely an incremental update. The hierarchical principle structure and adaptive critique loop represent qualitative changes in how we encode values into models. We believe this approach is more scalable to higher capability levels than any approach we know of that relies primarily on human feedback." — Anthropic Alignment Science Team, CAI v2 Technical Report
Why This Matters for Safety
The empirical results are encouraging. On Anthropic's internal harm evaluation suite — which covers 47 categories of potentially harmful output including misinformation, manipulation, weapons information, and hate speech — Claude 4 Opus trained with Constitutional AI v2 shows a 64% reduction in harmful output rate compared to an equivalent model trained with CAI v1. Critically, this improvement does not come at the cost of over-refusal: the rate at which the model incorrectly refuses benign requests actually dropped by 41%. Prior alignment approaches often traded one problem for the other.
Self-consistency is another key metric. A well-aligned model should hold consistent values across rephrased versions of the same question, across different conversation contexts, and across adversarial prompting attempts. CAI v2 improves self-consistency scores by 38% on Anthropic's values probes, which involve pairs of semantically equivalent prompts designed to elicit different responses from poorly aligned models. This is particularly relevant for jailbreak resistance — if a model's values are genuinely consistent rather than surface-level, adversarial rephrasing attacks are far less effective.
From a regulatory perspective, Constitutional AI v2 also offers something rare in the AI landscape: auditable values. The 78 principles in the constitution are published alongside the model card. Regulators, auditors, and enterprise procurement teams can inspect what the model is trained to value. This transparency is increasingly important as AI Act compliance requirements in Europe demand that high-risk AI systems demonstrate verifiable value alignment rather than simply asserting it. Anthropic's approach gives them a head start on every competitor that relies on opaque RLHF fine-tuning without any explicit principle set.
The full Constitutional AI v2 technical report is available on the Anthropic research page. The paper is dense and assumes familiarity with RLHF and preference learning, but for practitioners, Sections 4 and 6 — covering the hierarchical principle structure and the adaptive critique loop respectively — are the most significant contributions and well worth the time investment.
