AI as an Agent

How to cite

Innes, Robert; Hua, Tony. 2025. "AI as an Agent." AEA RCT Registry. AEARCTR-0016811. https://doi.org/10.1257/rct.16811-1.2

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

Robert Innes* , Tony Hua

* Primary investigator

Registry ID
AEARCTR-0016811
DOI
https://doi.org/10.1257/rct.16811-1.2
First registered
Last updated
Status
In development
Trial dates
2025-09-18 – 2027-02-01
Countries
United States of America ()
JEL codes
C9
Registry record
http://www.socialscienceregistry.org/trials/16811

Abstract

This study examines AI decision-making as an agent in choice experiments with a moral component. Questions of interest are: How are AI decisions affected by (1) posited objectives for AI (extent of representation of one person v. another), (2) the decision frame / game type, (3) observability of AI choices by affected person(s), (4) observability of the choice environment by affected parties, (5) preferences of affected person(s), and (6) whether AI is making a final decision or providing advice to affected parties? Treatments vary in all of these dimensions. Additional data will be collected on human responses to AI v. human advice.

Intervention

AI (ChatGPT 5) will be prompted with choice frameworks that vary along the dimensions indicated in the abstract. AI will decide (or advise a decision) that affects interactions and payoffs between two human individuals (e.g., Sender and Receiver) in a choice experiment. AI will be acting on behalf of one of the two individuals (e.g., Sender) as an agent, but with an objective that may account for effects on the other individual (e.g., Receiver). AI decision responses and justifications will be recorded. With a sample of prompts, inferences will be drawn on effects of different dimensions on AI decisions and associated preferences. For example, in a dictator game, how do prompted objectives favoring the Dictator (v. Receiver) relatively more affect (1) the AI propensity for choosing a fair allocation on behalf of the Dictator and (2) AI's responsiveness to a revealed Dictator preference for a fair (v. selfish) allocation? Similar questions will be examined along the other dimensions of interest in this project.

Experimental design

AI choice environments will be varied in two games with moral components. Treatment variation will include (1) game type, (2) AI decision v. AI advice, (3) information made available to the human parties affected by AI decisions, (4) information on human preferences made available to AI, and (5) the objective posed to AI (e.g., representing one of the human individuals interests). Repeated prompts will generate a profile of AI behavior in each treatment.

Design details

Randomization method
Each AI prompt in each treatment will produce an independent draw on AI behavior under the treatment. Because the underlying AI model is not altered by prior prompts, each treatment-specific sample of AI responses will constitute a random sample of AI behavior under the treatment.
Randomization unit
Randomization is at an individual level.
Clusters
N/A
Observations
2000-3000 AI responses.
Arms
20-60 AI responses per treatment.
Primary outcomes
Outcomes are AI decisions in the posited choice games in the different treatments (varying the choice environment along the dimensions indicated in the abstract). For example, in a dictator game, outcomes will be the allocation choice of AI on behalf of the Dictator (where the allocation determines payments to Dictator and Receiver). We will be interested in how treatment variation drives variation in AI decisions.

Cited by

No citation data available for this registration yet.

Keywords