AI Misinformation and Trust in News

How to cite

Sen, Ananya; Campante, Filipe; Durante, Ruben; Hagemeister, Felix. 2025. "AI Misinformation and Trust in News." AEA RCT Registry. AEARCTR-0015359. https://doi.org/10.1257/rct.15359-1.1

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

Ananya Sen* , Filipe Campante , Ruben Durante , Felix Hagemeister

* Primary investigator

Registry ID
AEARCTR-0015359
DOI
https://doi.org/10.1257/rct.15359-1.1
First registered
Last updated
Status
In development
Trial dates
2025-02-22 – 2025-04-22
Countries
Private
JEL codes
D91, L82
Registry record
http://www.socialscienceregistry.org/trials/15359

Abstract

This study aims to understand the extent to which technology-driven misinformation can alter institutional trust. In particular, we focus on how AI-generated misinformation can increase or decrease trust in media outlets. We partner with a large news outlet in Germany to test this by asking users to take a survey. These surveys are a standard part of the news outlet's way of doing consumer research. We randomly assign online users on the website into treatment and control groups within one such survey. In the treatment group, we show them three pairs of pictures sequentially and ask them to guess which one is AI-generated and which one is real. In the control group, we show them similar pairs of real pictures but ask them questions unrelated to AI. We then elicit their subjective beliefs about the importance of misinformation and their trust in our partner news outlet and other media platforms. We will also track their browsing and subscription behavior after the experiment.

Intervention

We randomly assign online users on the news website into treatment and control groups within one survey. In the treatment group, we show them three pairs of pictures sequentially and ask them to guess which one is AI-generated and which one is real. In the control group, we show them similar pairs of pictures but ask them questions unrelated to AI. We will swap out the AI generated picture with similar real ones in the control group.

Experimental design

We randomly assign online users on the news website into treatment and control groups within one survey. In the treatment group, we show them three pairs of pictures sequentially and ask them to guess which one is AI-generated and which one is real. In the control group, we show them similar pairs of pictures but ask them questions unrelated to AI. We will swap out the AI generated picture with similar real ones in the control group.

We will have a between-subject design such that an individual will be asked to complete a survey quiz only once and will be assigned to only one of the two groups. The main outcome variables of interest will be to look at subjective beliefs post the survey, browsing and clicking behavior on the website in the short and medium run.

Apart from these baseline treatment effects, we will also analyze heterogeneity on certain dimensions. We will analyze how the treatment effect varies by type of online visitor (such as subscriber/non-subscriber, heavy vs. light user) and how well the individual did on the quiz. A priori, the direction of the effect along these dimensions is ambiguous.

Design details

Randomization method
Randomization will be done with a computer software.
Randomization unit
Individual level
Clusters
--
Observations
20,000 -- this could be subject to some change based on survey uptake and traffic allocation by the company.
Arms
10,000 in treatment and 10,000 in control
Primary outcomes
(1) Subjective beliefs about trust in media and the importance of misinformation (2) Preference for subscribing to the news outlet. (3) Longer-term browsing and subscription behavior

Cited by

No citation data available for this registration yet.

Keywords