The Conformity Effects of AI-Generated Advice on Individual Choice

How to cite

Nakamoto, Yasuhiro; Kitadai, Ayato. 2026. "The Conformity Effects of AI-Generated Advice on Individual Choice." AEA RCT Registry. AEARCTR-0017799. https://doi.org/10.1257/rct.17799-1.0

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

Yasuhiro Nakamoto* , Ayato Kitadai

* Primary investigator

Registry ID
AEARCTR-0017799
DOI
https://doi.org/10.1257/rct.17799-1.0
First registered
Last updated
Status
In development
Trial dates
2026-02-02 – 2026-02-28
Countries
Private
JEL codes
C92, D03
Registry record
http://www.socialscienceregistry.org/trials/17799

Abstract

This study examines how advice from artificial intelligence influences human image-recognition decisions. We conduct a 2×2 experiment that varies (i) the source of advice (human vs. AI) and (ii) the number of advisors (single vs. multiple). Participants first make an individual image-recognition judgment and then receive advice generated either by one human, multiple humans, one AI system, or multiple AI systems. We test whether and how participants revise their decisions in response to these different advice conditions. We use four images that are known to be prone to AI hallucination-like errors, which allows us to study advice effects under challenging image-recognition tasks. We plan to conduct both a laboratory experiment and a web-based (online) experiment.

Intervention

Participants view images (e.g., pictures containing apples and other objects) for approximately 10 seconds. After each image, they are asked to choose, from multiple-choice options, how many target objects (e.g., apples) were shown. Participants then receive advice about the number of objects, provided either by a human advisor or an AI advisor.

Experimental design

We conduct a 2×2 between-subjects experiment that varies (i) the source of advice (human vs. AI) and (ii) the number of advisors (single vs. multiple). Participants are randomly assigned to one of four conditions. In each task, participants view an image for approximately 10 seconds and then select, from multiple-choice options, the number of target objects shown. After making an initial judgment, participants receive advice corresponding to their assigned condition and then provide a final answer. The primary outcome is whether participants revise their initial choice after receiving advice.

Design details

Randomization method
Participants are randomly assigned to one of the four treatment arms by the research team in the office.
Randomization unit
Session-level assignment: the experimental condition is set for each session (all participants in the same session receive the same condition). Within each session, the order of the four images is randomized for each participant by the experiment software.
Clusters
about 12 laboratory sessions (about 20 participants per session; total N=240)
Observations
240 participants in the laboratory experiment and 200–500 participants in the web-based (online) experiment.
Arms
Laboratory experiment: 50 participants per arm (4 arms; total N=240). Web-based experiment: 50–100 participants per arm (4 arms; total N=200–400).
Primary outcomes
Choice revision (binary): whether a participant changes their initially selected answer option after receiving advice (AI vs. human; single vs. multiple advisors).

Cited by

No citation data available for this registration yet.

Keywords