Skip to content

Paper 042 / Anthropic

The Capacity for Moral Self-Correction in Large Language Models

The paper studies whether instruction-following models can reduce biased responses when prompts explicitly ask them to consider fairness or avoid stereotypes.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / alignment_and_safety

The paper studies whether instruction-following models can reduce biased responses when prompts explicitly ask them to consider fairness or avoid stereotypes.

CONTRIBUTION / 01

Contribution 1

Evaluated moral self-correction across several social-bias benchmarks.

CONTRIBUTION / 02

Contribution 2

Compared instruction strength, model scale, and training regimes.

CONTRIBUTION / 03

Contribution 3

Identified settings where prompting reduces bias and settings where gaps remain.

02

Research context

alignment_and_safety / 2023

The Capacity for Moral Self-Correction in Large Language Models places bias inside the broader alignment and safety discussion at Anthropic, with self correction supplying a second analytical lens. Read together, the source records three advances: Evaluated moral self-correction across several social-bias benchmarks; Compared instruction strength, model scale, and training regimes; and Identified settings where prompting reduces bias and settings where gaps remain. Keeping those moves together prevents fairness from being detached from its evidence. For an implementation review, the relevant consequence is that behavioral reminders can reduce some measured biases, but prompts do not replace dataset, model, and outcome audits.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in The Capacity for Moral Self-Correction in Large Language Models tests evaluated moral self-correction across several social-bias benchmarks and compared instruction strength, model scale, and training regimes against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

The evidentiary value of The Capacity for Moral Self-Correction in Large Language Models comes from the relationship among its reported moves. Two entries define the method-level claim: Evaluated moral self-correction across several social-bias benchmarks; and Compared instruction strength, model scale, and training regimes. The cataloged result is: Identified settings where prompting reduces bias and settings where gaps remain. On that basis, behavioral reminders can reduce some measured biases, but prompts do not replace dataset, model, and outcome audits. The catalog nevertheless records that the demonstrated bias result belongs to a setup defined by model revisions, selected threat model, evaluator models, construct validity, rater instructions, and prompt sampling, not to every later self correction system. Reproduction work should separate genuine bias transfer from behavior caused by a changed self correction setup. A reproduction ledger for The Capacity for Moral Self-Correction in Large Language Models should preserve bias, vary self correction, and retain a counterexample tied to Compared instruction strength, model scale, and training regimes before judging transfer.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in The Capacity for Moral Self-Correction in Large Language Models supports identified settings where prompting reduces bias and settings where gaps remain.
05

Practical implication for AI builders

Anthropic / 2023

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / bias

For a proposed BrokenGPT experiment based on The Capacity for Moral Self-Correction in Large Language Models, pair fairness-oriented system prompts with benchmark audits and production sampling, recording where prompt-based correction fails by domain and language. Keep the bias path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / self correction

Validate the proposed bias route against the paper's reported outcome: Identified settings where prompting reduces bias and settings where gaps remain. A proposed The Capacity for Moral Self-Correction in Large Language Models gate needs refusal precision, adversarial coverage, and helpful-answer retention; its self correction cases should remain disaggregated from the overall bias score.

TRADEOFF / 03

Proposed decision boundary / fairness

Balance usefulness, oversight burden, and residual risk before promoting the proposed fairness design. Because operational use of bias introduces unsampled behaviors, language coverage, adversarial adaptation, judge bias, and deployment drift, so a matched replay is necessary before a release decision, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • The demonstrated bias result belongs to a setup defined by model revisions, selected threat model, evaluator models, construct validity, rater instructions, and prompt sampling, not to every later self correction system.
  • Operational use of bias introduces unsampled behaviors, language coverage, adversarial adaptation, judge bias, and deployment drift, so a matched replay is necessary before a release decision.

PRIMARY SOURCES

  1. 01
    The Capacity for Moral Self-Correction in Large Language Models

    Anthropic — Primary primary arXiv paper / 15 February 2023 / Deep Ganguli, Amanda Askell, Nicholas Schiefer, and 46 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does The Capacity for Moral Self-Correction in Large Language Models study?

The paper studies whether instruction-following models can reduce biased responses when prompts explicitly ask them to consider fairness or avoid stereotypes.

02Which methods does The Capacity for Moral Self-Correction in Large Language Models use?

The experimental design in The Capacity for Moral Self-Correction in Large Language Models tests evaluated moral self-correction across several social-bias benchmarks and compared instruction strength, model scale, and training regimes against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does The Capacity for Moral Self-Correction in Large Language Models report?

The reported evidence in The Capacity for Moral Self-Correction in Large Language Models supports identified settings where prompting reduces bias and settings where gaps remain.

04What is the proposed BrokenGPT application for The Capacity for Moral Self-Correction in Large Language Models?

Proposed: pair fairness-oriented system prompts with benchmark audits and production sampling, recording where prompt-based correction fails by domain and language.

MAJOR LAB RESEARCH / PAPER 042

Continue after The Capacity for Moral Self-Correction in Large Language Models

After The Capacity for Moral Self-Correction in Large Language Models, browse the full index for adjacent alignment_and_safety research and work from Anthropic.

Open the paper index