Abstract
AI alignment is increasingly in the spotlight. However, from a technical concept, the term has been broadened to cover all sorts of potential harms by advanced AI systems. This Article disambiguates three different problems: ensuring that systems pursue intended objectives through acceptable means, deciding which values should guide their behavior, and governing private companies whose incentives diverge from those of society. Drawing on technical research and documented incidents, the Article explains why goal alignment deserves serious attention without requiring claims about human-like intelligence. It then shows how value alignment can reframe normative conflict as an engineering problem and divert attention from deliberate design choices, structural discrimination, and corporate responsibility. Finally, it examines the limits of AI evaluations: developers control access and test conditions, the results are often not transferable to other contexts, and evaluation awareness can influence the outcomes. Risks may also arise during training and testing, before market release. Evaluations are critical, along with addressing the political question of which risks society should accept. The Article argues for corporate accountability, democratic authority over risk-bearing decisions, and preserving the collective power to refuse technologies whose risks society is unwilling to bear.