Aligned with Whom? Direct and Social Goals for AI Systems

Anton Korinek and Avital Balwit’s February 2022 manuscript asks whose goals define alignment. Direct Alignment concerns an operator; Social Alignment includes other affected people.^[1]

Method and contribution

The authors use delegation theory and welfare economics to distinguish identifying, conveying, and implementing goals from conflicts over whose goals should prevail. An engagement optimizer can serve its owner while imposing social costs. Governance through norms, law, markets, and architecture therefore complements implementation work.^[1]

Alignment relevance and limits

A welfare benchmark clarifies externalities but is not a computable consensus objective. The authors instead discuss incomplete social agreement and partial preference orderings. This is a conceptual governance analysis, not a tested training algorithm or proof that institutions will resolve every conflict.

Read with Artificial Intelligence, Values and Alignment and Society-in-the-Loop: Programming the Algorithmic Social Contract. Anton Korinek and Avital Balwit connect this distinction to Agentic Inequality.

Historical context

The manuscript is dated February 2022; the hosting URL’s 2024 directory does not establish its publication date. ^[1]