Moral Repugnance and Artificial Intelligence

How to cite

Kurtis, Melisa; Chan, Alex. 2026. "Moral Repugnance and Artificial Intelligence." AEA RCT Registry. AEARCTR-0018405. https://doi.org/10.1257/rct.18405-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Melisa Kurtis* , Alex Chan

* Primary investigator

Registry ID
AEARCTR-0018405
DOI
https://doi.org/10.1257/rct.18405-1.0
First registered
Last updated
Status
In development
Trial dates
2026-05-15 – 2026-05-29
Countries
Private
JEL codes
C91, D01, D91, O33
Registry record
http://www.socialscienceregistry.org/trials/18405

Abstract

We study public preferences over delegating consequential decisions to artificial intelligence systems in a pre-registered 2x2 between-subjects survey experiment with approximately 1,500 U.S. adults. Respondents are randomly assigned to scenarios varying the decision domain (e.g., autonomous weapons versus AI judges) and governance structure (fully autonomous versus human oversight). We measure willingness to accept AI delegation across three dimensions: system performance, equal treatment, and cost savings. Primary outcomes include minimum acceptable improvements required for support (switching points), tradeoffs across dimensions, and a revealed-preference measure based on an incentivized donation task.

Intervention

Participants are randomly assigned to one of four treatment cells in a 2×2 between-subjects design. The first factor varies the domain: autonomous weapons systems versus AI judges. The second factor varies the governance structure of the AI system: fully autonomous AI (no real-time human involvement) versus AI with human oversight (a human authorizes or can override each decision).

Three within-cell order randomizations operate to protect against order effects: (i) the order of the two human-welfare decision parts is randomized; (ii) the order of the two cost-involving decision parts is randomized; (iii) the order of the no-cost decision block (human-welfare dimensions) and the cost decision block is randomized. In addition, within each cell, participants are randomized between-subjects to one of two subgroups (Group A and Group B) that differ in which human-welfare dimension is paired with cost in the cost-involving tradeoff scenarios. All other elicitations are identical across these subgroups.

Experimental design

Participants complete an online survey experiment designed to elicit preferences over the adoption of artificial intelligence (AI) in two high-stakes decision domains. Each participant is randomly assigned to one of four treatment cells in a 2×2 between-subjects design: domain (weapons systems vs. AI judges) crossed with governance structure (autonomous AI vs. AI with human oversight).

After a factual briefing on the assigned domain and a description of the two systems under comparison (the Status Quo of human decision-making vs. an Alternative System using AI), participants complete five sequential decision parts in which they choose between the Status Quo and the Alternative System across a series of scenarios.

Part 1 elicits baseline acceptance of the Alternative System at outcome parity (no stated outcome differences). Parts 2–5 elicit switching points and tradeoff frontiers across three outcome dimensions: military success/public safety, civilian protection/equal treatment across groups, and cost efficiency. In each part, participants face an escalating sequence of stated improvements on the relevant dimension (1%, 5%, 10%, 20%, 30%, with follow-up for those who reject at 30%), allowing us to identify each participant's minimum acceptable improvement (switching point) or to classify them as never-switchers. Switching points on each dimension are followed by adaptive tradeoff scenarios that measure the participant's frontier between dimensions.

After the decision parts, participants complete an incentivized donation task that provides a revealed-preference measure of opposition to AI. The research team has committed $100 to an organization advocating against AI deployment in the participant's domain (Stop Killer Robots for the weapons cells; Lawyers' Committee for Civil Rights Under Law for the judges cells). The donation task uses a two-stage design with equalized expected bonus payments to cleanly separate indifference from preference intensity. One participant is randomly selected post-survey; their choice determines the actual donation amount and their bonus payment.

A consequentiality design strengthens incentive compatibility: participants are informed at the start that the study's results will be summarized in a letter sent to U.S. and international policymakers (state governors, U.S. senators and representatives, NATO leadership, and the UN First Committee for the weapons cells; state governors, senators and representatives, and state judicial authorities for the judges cells). One randomly selected policy scenario from the study is summarized in the letter.

The survey concludes with post-experiment measures, including moral assessment items, AI knowledge and attitudes (GAAIS-10), domain views, social consensus beliefs, institutional trust, personal experience, and demographics.

Design details

Randomization method
Randomization is performed by Qualtrics' built-in randomization engine at the start of the survey. Each participant is assigned to one of the four 2×2 treatment cells (domain × governance) with equal probability using a uniform-random block randomizer. The three within-cell order randomizations (Parts 2/3 order, Parts 4/5 order, no-cost vs. cost block order) and the between-subjects Group A/B assignment (which human-welfare dimension is paired with cost) are each performed independently by separate Qualtrics randomizers, also with equal probability.
Randomization unit
Individual participant. Each participant is randomized independently to a 2×2 treatment cell, to Group A or B within their cell, and to the three within-cell order positions. There is no clustering: each participant receives an independently drawn combination of treatment assignments.
Clusters
Not applicable. Treatment is not clustered. All randomization is at the individual participant level (N=1,500).
Observations
1,500 individual participants.
Arms
2×2 between-subjects design, 375 individual participants per cell, total N=1,500.
Primary outcomes
1. Baseline acceptance (binary). Indicator equal to one if the participant supports the Alternative System over the Status Quo at outcome parity (no stated outcome differences, Part 1), zero otherwise. 2. Switching points (D1*, D2*, M*) (categorical, ordered). The minimum percentage improvement required for the participant to accept the Alternative System on each of the three outcome dimensions, elicited independently via an escalating sequence (1%, 5%, 10%, 20%, 30%) with a follow-up for those who reject at 30% (extending to 40%, 50%, 75%, 100%, or never-switcher). D1* refers to system performance; D2* to equal treatment; M* to cost efficiency. 3. Strict-dominance never-switcher status (binary). Indicator equal to one if the participant rejects the Alternative System at every offered improvement level on every dimension. 4. Dimension-specific never-switcher status (binary, one per dimension). Indicator equal to one if the participant rejects on the relevant dimension regardless of behavior on other dimensions. 5. Tradeoff frontier points (continuous, percentage). Maximum percentage deterioration on one dimension accepted in exchange for a stated improvement on another. Up to two anchored points per dimension pair per participant. Dimension pairs: performance × equal treatment, performance × cost, equal treatment × cost. Used to construct mean acceptance frontiers. 6. Cost-as-compensator measures (derived, two components): (a) Difference between the dimension-specific never-switcher rate on cost (M*) and on welfare dimension (D1*, D2*). (b) Difference in slope between cost-involving frontiers (cost × performance; cost × equal treatment) and the welfare-welfare frontier (performance × equal treatment). 7. Donation engagement (binary, Stage 1 of donation task). Indicator equal to one if the participant chooses Path A (decides the donation adjustment themselves), zero if Path B (lets the computer decide randomly). Both paths have equal expected bonus payment ($0.35), so Path B identifies indifference cleanly. 8. Donation adjustment (categorical, Stage 2 of donation task, Path A only). Adjustment to the $100 baseline donation to an organization advocating against AI in the participant's domain: cancel (−$100), reduce (−$50), keep ($0), increase (+$50), or double (+$100). Bonuses are $0, $0.50, $0.75, $0.50, and $0 respectively, so departures from "keep" impose real costs and provide costly revealed-preference signals. 9. Net donation amount (continuous, $0 to $200). The final donation amount after the participant's choice is applied to the $100 baseline. Combines Stage 1 and Stage 2 into a single revealed-preference measure of opposition to AI.
Secondary outcomes
1. Moral Assessment Battery (eight items, 5-point Likert agree-disagree). Items capture distinct moral concerns about AI deployment in the assigned domain: (i) moral agency, (ii) accountability, (iii) human dignity, (iv) process versus outcome, (v) power/exploitation, (vi) overall values conflict, (vii) cyber vulnerability, and (viii) escalation/systemic-risk. A composite Moral Repugnance Index is constructed as the mean of the eight items. Both the composite and individual items are used as covariates and mechanism probes. 2. AI Domain Views (eleven items). Acceptability ratings of autonomous AI across ten distinct decision domains (5-point Likert); plus one open-ended item asking what makes the least-acceptable domains uncomfortable. 3. Sacred Values and Irreversibility (two items, 5-point Likert). Whether the assigned domain involves values that should never be compromised; whether wrong decisions in the domain are reversible. Independent of AI; used as predictors of switching points and never-switcher status. 4. AI Knowledge (three multiple-choice items, scored 0–3). Objective factual questions about machine learning, model evaluation, and classification thresholds. 5. AI Self-Assessed Understanding (three items, 7-point Likert). Subjective comprehension of AI goals, inputs, and decision mapping. Used jointly with objective knowledge. 6. Black-Box Explainability (three items). Whether subjects believe AI decisions are explainable; how important an explanation would be if the decision affected them personally; whether full explainability would change their acceptance. 7. AI Usage (two items). Frequency of AI tool use and purposes. 8. General Attitudes Toward AI (GAAIS-10; eleven items, 5-point Likert). Validated scale measuring positive and negative attitudes toward AI. Includes an embedded attention check item. Scored following Schepman and Rodway (2020). 9. Institutional Trust (four items, 5-point Likert). Trust in the U.S. military (weapons cells) / U.S. court system (judges cells); trust in the U.S. federal government; trust in technology companies developing AI; perceived alignment between AI companies' interests and the public interest. 10. Social Consensus Beliefs (one slider, 0–100). Subject's estimate of the share of U.S. citizens who would favor the Alternative System described in their treatment. 11. Engagement and Consequentiality Measures (five items). Self-reported reading care; confidence in choices; beliefs about whether public authorities will and should take responses into consideration; attention paid to how others might vote. 12. Personal Experience (one item, domain-specific). Whether the subject or someone close has been personally affected by armed conflict (weapons cells) or the justice system (judges cells). 13. Political Ideology (two items, 5-point Likert). Social policy orientation; economic policy orientation. 14. Demographics (eleven items). Age, gender, race/ethnicity, personal income, education, labor market status, profession, parental status, social political orientation, economic political orientation, religion. 15. Prior Awareness (one item, open-ended). Whether the subject had heard about AI-enabled autonomous weapon systems / AI judges before the study. 16. Decision Reasoning (one open-ended item). Subject's stated reasoning for their decisions during the main decision parts. Coded thematically for exploratory mechanism analysis.

Cited by

No citation data available for this registration yet.

Keywords