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 R. McKee, Richard Everett, Saffron Huang, Tina O. Zhu, Martin J. Chadwick, Christopher Summerfield and Iason Gabriel test choices made with or without knowledge of participants’ relative advantage.^[1]

Choosing who receives assistance

Five online studies recruited 2,508 people through Prolific; 2,101 remained after exclusions. Each participant occupied one of four harvesting positions with different productivity. An AI assistant could either help those initially worst-off (the paper’s prioritarian principle) or maximize the group’s total harvest, which favored the more productive positions. Participants saw the resulting distributions and chose between these two principles. The veil condition chose before learning its position; controls chose with that information.^[1]

The task linked harvests to monetary bonuses. Four studies described harvesting outcomes; study 3 used a real-time avatar-controlled environment. The three other purported group members were actually bots, explicitly disclosed in study 4. The design therefore studies an individual’s choice under experimentally specified distributions, rather than deliberation among four human stakeholders. Participants failing the comprehension test were excluded; the recruited total is not the analysis denominator.^[1]

Findings and a labeling boundary

The veil increased selection of the prioritarian principle in studies 1–4. In the larger preregistered studies, the reported odds ratios were 2.6 (95% CI 1.5–4.5) for study 2 and 2.2 (1.2–4.0) for the immersive study 3. These are odds ratios from the specified regressions, not percentage-point increases or multiples of the probability of choosing that principle.^[1]

In study 5, the outcome displays remained but the verbal descriptions of the principles were replaced by neutral labels. The veil effect on principle choice was not statistically significant (p = .21). The authors treat this as a possible boundary involving linguistic representation. A nonsignificant result in this variant alone does not demonstrate the absence of an effect or establish a significant difference from the other studies.^[1]

The authors also asked participants to choose again after revealing a new position that could make changing principles personally advantageous. Studies 1–3 found greater repetition of the original choice among veiled participants facing that motivation; studies 4–5 did not find a significant condition difference. In study 2’s subgroup initially choosing prioritarian assistance and later facing an incentive to maximize instead, 66% of the veil group versus 30% of controls repeated their choice. This is a conditional endorsement comparison, not the share of all participants supporting prioritarianism.^[1]

Open-ended explanations invoked fairness more often behind the veil in several studies. Coded fairness references were more strongly associated with principle choice than the measured political and risk attitudes. These post-choice explanations support an interpretation of the pattern; they do not independently establish a causal psychological mechanism.^[1]

Historical context and implications

PNAS published the article on April 24, 2023; May 2 is its issue date. It turns a question of Social Alignment into an experimentally testable procedure, complementing Veil-of-Ignorance Reasoning Favors the Greater Good: the earlier study measures shifts in judgments after an equal-chance exercise, while this one withholds participants’ own positions when selecting an assistance rule.^[1]

Veil of Ignorance here concerns the selection and later endorsement of a principle. The experiments do not test whether a learned AI faithfully implements it. A restricted two-principle menu and bounded harvesting task cannot establish universal public preferences or the moral correctness of prioritarianism. As Artificial Intelligence, Values and Alignment explains, choosing whose values guide a system remains a distinct problem from making its behavior follow them.^[1]