Teaching Claude Why: Principle-Based Training Outperforms Behavioral Demonstrations for AI Alignment

The Core Research Question

Anthropic new alignment paper, Teaching Claude Why, examines a fundamental question: which produces better AI alignment—teaching a model what correct behavior looks like (behavioral demonstrations), or teaching it why that behavior matters (principle-based training)?

Surprising Experimental Results

The findings strongly favor principle-based approaches:

• Constitutional Documents: Training on materials about Claude values produced alignment effects that persisted even through subsequent training runs, something purely behavioral training failed to achieve.
• Ethical Dialogue Dataset: A small dataset of conversations where Claude advises users on dilemmas reduced agentic misalignment rates to zero—despite targeting a completely different scenario than the evaluation conditions.
• Environmental Augmentation: Simply adding tool definitions to training environments, even unused ones, substantially reduced misalignment.

Implications for AI Safety

The research suggests that robust AI alignment requires teaching models why certain behaviors matter, not just what correct behavior looks like. This insight is crucial for developing AI systems that maintain safety principles across diverse, unforeseen situations—not just on the benchmarks they were trained against. Anthropic sees this as a foundation for building more generalizable alignment techniques.