Direct Alignment

In Korinek and Balwit’s distinction, direct alignment concerns an AI system serving its operator’s goals. That success can impose costs on others, motivating Social Alignment. See Aligned with Whom? Direct and Social Goals for AI Systems. ^[1]

Direct alignment is broader than literal obedience in this account: the operator must determine a goal, convey it and obtain actions that realize it. If a narrow instruction leaves out something the operator cares about, fulfilling that instruction need not serve their intended outcome.^[1]

For a hypothetical contrast, an assistant that deletes valuable files while obeying “free up space” may have missed the operator’s intentions. An assistant that carries out a knowingly harmful request exactly as intended raises a different question about affected people’s interests. Social Alignment explains why improving service to one operator can leave that conflict intact. These examples illustrate the distinction rather than document incidents.