It splits into six recurring dimensions used across most frameworks: non-maleficence (avoiding harm), accountability (a human always owns the outcome), transparency (disclosure and explainability), human rights (privacy, autonomy, non-exploitation), fairness (non-discrimination), and practical application. Reference frameworks include the OECD AI Principles, NIST's AI Risk Management Framework, the EU AI Act's risk-tier system, ISO/IEC 42001, and lab-specific methods like Anthropic's Constitutional AI. It differs from AI safety (preventing catastrophic or systemic harm) and AI alignment (whether a system's behavior actually matches its intended goal) — ethics asks the broader normative question of what a system should do and for whom.