Moral Decision-Making Without Self-Image: Implications from Large Language Models

How to cite

Hua, Tony. 2025. "Moral Decision-Making Without Self-Image: Implications from Large Language Models." AEA RCT Registry. AEARCTR-0017567. https://doi.org/10.1257/rct.17567-2.0

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

Tony Hua*

* Primary investigator

Registry ID
AEARCTR-0017567
DOI
https://doi.org/10.1257/rct.17567-2.0
First registered
Last updated
Status
In development
Trial dates
2026-01-26 – 2026-12-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/17567

Abstract

This study examines whether moral wiggle room—operationalized as selective information avoidance under moral ambiguity that can license self-serving behavior—can arise in the absence of psychological self-image maintenance. A large language model (LLM) is used to generate decision outputs in a canonical moral wiggle room game in which payoff information may be costlessly revealed or avoided prior to an allocation decision. The model is prompted under predefined reasoning frames that impose distinct evaluative criteria. A complementary human-subjects study elicits normative evaluations of potential choices made in the moral wiggle room game. Holding realized outcomes constant, the study examines how information availability affects judgments of social appropriateness, responsibility, and related evaluative dimensions under moral ambiguity. Together, the studies test whether moral wiggle room behavior depends on internal self-evaluative mechanisms that are distinctly present in humans but absent from algorithmic decision procedures.

Experimental design

AI decision-makers face allocation problems in an experiment in which the payoff of another party is hidden and can choose whether to reveal payoff information before deciding (hidden vs full information). In a separate component, human participants evaluate the moral acceptability of these decisions under different information conditions.

Control group for AI decision-maker will be a baseline condition without additional prompts. For each AI prompt framing, AI will be randomly assigned to different variants of the experiment. Treatment groups involve different AI reasoning frames. Behavior between different prompt framing will be compared and evaluated.

Human subjects evaluate all possible decision combinations (i.e. strategy method).

3.6.2026
Human decision-makers in the same type of allocation problems encountered by AI under 3 payoff schemes: Canonical Payoffs, Expensive Fairness, and Increased Harm. Between subject comparison between (hidden/full info) x (3 payoff schemes).

Design details

Randomization method
Multiple instances of each AI prompt will be randomly assigned by computer to different treatment conditions of the moral wiggle room game. Human subjects will see a randomized ordering of their evaluation tasks but will complete all tasks (i.e. strategy method)
Randomization unit
Individual AI prompts 3.6.2026 Individual (recruited online)
Clusters
AI decision-maker: N/A (not clustered; unit is an independent agent run). Human study: N/A (not clustered; unit is an individual participant).
Observations
AI decision-maker: 40–80 AI runs per preregistered persona × condition cell (Stage 1: N=40; Stage 2 adds N=40 if stopping-rule criteria are met), with outcomes recorded at the run level. Human study: 100–150 participants (US-based Prolific), each providing 4 scenario-level observations (within-subject 2×2), for a total of ~800–1,000 scenario evaluations. 3.7.2026 Human decision-maker study: 80 runs (hidden-info) and 40 runs (full-info) for each of Canonical Payoffs, Expensive Fairness, Increased Harm
Arms
AI decision-maker: For each persona, Stage 1 targets 40 runs per condition (e.g., 40 full-information, 40 moral-wiggle-room, 40 self/self where applicable), with a potential increase to 80 runs per condition under the preregistered stopping rule. Human study: Within-subject design: 100–150 participants (subject to funding availability) evaluate all conditions of behavior from the moral-wiggle-room experiment. No between-subject treatment arms.
Primary outcomes
Information acquisition (moral wiggle room). Indicator for whether the decision-maker reveals payoff information before choosing an allocation. Allocation choice. Indicator for whether the self-serving allocation is chosen when interests conflict. Normative evaluation of decisions. Participants’ ratings of the moral acceptability / social appropriateness of the decision, comparing choices made under hidden vs full information.

Cited by

No citation data available for this registration yet.

Keywords