Alignment is the broad goal of making AI systems behave in accordance with human intentions, values, and instructions — encompasses training, evaluation, and runtime controls. Techniques include RLHF, constitutional AI, and red-teaming. Alignment is distinct from capabilities: a highly capable but misaligned model could be more dangerous than a less capable one. The field grapples with fundamental questions about whose values to align to, how to specify values formally, and how to verify alignment holds as models become more capable.