Using the Veil of Ignorance to align AI systems with principles of justice

Can an impartial choice procedure help people select principles for AI? Laura Weidinger, Kevin McKee, Iason Gabriel and coauthors test choices made with or without knowledge of participants’ relative advantage.^[1]

Method and findings

Five incentivized online studies recruited 2,508 people, with 2,101 retained after exclusions. Participants selected between assistance prioritizing the worst-off and assistance maximizing total returns in a harvesting task. Veiled participants more often chose the prioritarian principle in several conditions; removing verbal principle labels produced a boundary condition.^[1]

Alignment relevance and limits

Veil of Ignorance turns a question of Social Alignment into an experimentally testable procedure. The paper examines principle selection and subsequent endorsement, not whether an AI faithfully implements those principles.

A bounded task with a restricted principle menu cannot settle the moral correctness of prioritarianism or establish universal public preferences. The results inform normative debate rather than resolve it. See Artificial Intelligence, Values and Alignment for the broader framing and Using the Veil of Ignorance to align AI systems with principles of justice.

Historical context

PNAS published Using the Veil of Ignorance to align AI systems with principles of justice on April 24, 2023. The article record separately gives May 2 as the issue date, January 28 as acceptance, and August 25, 2022 as receipt; this event uses publication.^[1]

Laura Weidinger, Kevin McKee, Iason Gabriel and coauthors studied choices behind a Veil of Ignorance.