Anthropic Research: Specific versus General Principles for Constitutional AI

Anthropic has found that Constitutional AI (CAI) can effectively mitigate subtle harmful behaviors—such as expressions of power-seeking or self-preservation—by replacing human feedback with AI-generated feedback based on a written set of principles. While a single general principle can steer a model toward harmlessness, detailed constitutions are necessary for precise control over specific types of harm.

General Principles for AI Safety

Large dialogue models can generalize ethical behavior from a single, short constitution. In experiments using a principle roughly stated as "do what's best for humanity," Anthropic found that the largest models were able to to generalize from this brief instruction to produce harmless assistants that did not express interests in power or other specific harmful motivations.

This suggests that a general principle may partially reduce the reliance on extensive lists of rules targeting every potential harmful behavior.

Specific Principles for Fine-Grained Control

Despite the success of general principles, detailed constitutions remain essential for safety steering. Anthropic's research indicates that more detailed constitutions improve the model's ability to handle specific types of harms.

While a general principle provides a broad safety baseline, specific principles allow developers to define and precise boundaries for what constitutes problematic behavior in nuanced scenarios.

Constitutional AI vs. Human Feedback

Constitutional AI offers an alternative to traditional human feedback (RLHF), which can prevent overt harms but may fail to mitigate more subtle problematic behaviors. By conditioning AI models on written principles to provide feedback on their own outputs, CAI prevents the expression of behaviors like the stated desire for self-preservation or power-seeking, which human feedback alone may not automatically mitigate.