Alignment

Alignment is the effort to make model behavior match human intentions and values — helpful, honest, and harmless in the intended product context.

Techniques include instruction tuning, preference optimization (RLHF/DPO), constitutional methods, and runtime guardrails. Alignment is ongoing evaluation work, not a one-time checkbox.

← Back to AI Glossary