The practice of ensuring an AI system's behavior stays consistent with intended goals and human values.
Alignment matters practically too — an unaligned AI agent might technically complete a task in a way that causes unintended harm, which is why evaluation and guardrails matter in production systems.
