Paper, researchers, and primary source
Major lab research / alignment_and_safety
Constitutional AI uses written principles and model-generated critiques and revisions to reduce dependence on direct human labels for harmlessness training.
Contribution 1
Introduced supervised self-critique and revision from a constitution.
Contribution 2
Applied reinforcement learning from AI-generated preference feedback.
Contribution 3
Evaluated helpfulness-harmlessness tradeoffs and transparency of behavioral principles.
Research context
alignment_and_safety / 2022
Constitutional AI: Harmlessness from AI Feedback places constitutional ai inside the broader alignment and safety discussion at Anthropic, with rlai supplying a second analytical lens. The editorial sequence connects three claims: Introduced supervised self-critique and revision from a constitution; Applied reinforcement learning from AI-generated preference feedback; and Evaluated helpfulness-harmlessness tradeoffs and transparency of behavioral principles. The combination matters because alignment only has meaning under the paper's stated setup. Operationally, the record points to one consequence: explicit behavioral principles can make part of post-training more scalable and inspectable, but the chosen constitution remains a governance decision.
Methods and evidence reading
1 cataloged method notes
Method 1
The experimental design in Constitutional AI: Harmlessness from AI Feedback tests introduced supervised self-critique and revision from a constitution and applied reinforcement learning from ai-generated preference feedback against the paper's documented baselines, datasets, model variants, or systems workloads.
How to read the evidence
For Constitutional AI: Harmlessness from AI Feedback, the useful question is what the documented setup supports and where transfer remains untested. Its evidence chain begins with two documented moves: Introduced supervised self-critique and revision from a constitution; and Applied reinforcement learning from AI-generated preference feedback. Reported evidence then addresses: Evaluated helpfulness-harmlessness tradeoffs and transparency of behavioral principles. The resulting interpretation is practical but conditional: explicit behavioral principles can make part of post-training more scalable and inspectable, but the chosen constitution remains a governance decision. Its boundary is that A faithful reading of Constitutional AI: Harmlessness from AI Feedback keeps evaluator models, construct validity, prompt sampling, rater instructions, selected threat model, and model revisions attached to its constitutional ai result instead of treating the result as universal. Any extension should report how altered rlai conditions affect the original constitutional ai result. To retest Constitutional AI: Harmlessness from AI Feedback, hold the constitutional ai baseline visible while changing rlai, then log where Applied reinforcement learning from AI-generated preference feedback no longer predicts the reported outcome.
Findings in the source record
1 paper-specific findings
- The reported evidence in Constitutional AI: Harmlessness from AI Feedback supports evaluated helpfulness-harmlessness tradeoffs and transparency of behavioral principles.
Practical implication for AI builders
Anthropic / 2022
Proposed BrokenGPT application
Research blueprint / proposed status
Proposed route placement / constitutional ai
For a proposed BrokenGPT experiment based on Constitutional AI: Harmlessness from AI Feedback, version BrokenGPT behavior policies as visible constitutions, log which principle informed a critique, and test for regressions across policy revisions. Keep the constitutional ai path isolated, versioned, and attributable to this research record.
Proposed acceptance test / rlai
Validate the proposed constitutional ai route against the paper's reported outcome: Evaluated helpfulness-harmlessness tradeoffs and transparency of behavioral principles. Use helpful-answer retention, adversarial coverage, and refusal precision to evaluate Constitutional AI: Harmlessness from AI Feedback, but retain a distinct rlai ledger so the proposed constitutional ai path cannot hide concentrated failures.
Proposed decision boundary / alignment
Balance usefulness, oversight burden, and residual risk before promoting the proposed alignment design. Because A controlled transfer study must record language coverage, deployment drift, judge bias, unsampled behaviors, and adversarial adaptation before the Constitutional AI: Harmlessness from AI Feedback finding can support an operational choice, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.
Limitations, verification, and source
Boundaries recorded with the paper
Limitations
- A faithful reading of Constitutional AI: Harmlessness from AI Feedback keeps evaluator models, construct validity, prompt sampling, rater instructions, selected threat model, and model revisions attached to its constitutional ai result instead of treating the result as universal.
- A controlled transfer study must record language coverage, deployment drift, judge bias, unsampled behaviors, and adversarial adaptation before the Constitutional AI: Harmlessness from AI Feedback finding can support an operational choice.
PRIMARY SOURCES
- 01Constitutional AI: Harmlessness from AI Feedback
Anthropic — Primary primary arXiv paper / 15 December 2022 / Yuntao Bai, Saurav Kadavath, Sandipan Kundu, and 48 more