|
Field
Trial Status
|
Before
in_development
|
After
completed
|
|
Field
Abstract
|
Before
This study aims to understand the impact of providing information about the potential of AI to generate misinformation (deepfakes) on news consumption and subscription outcomes. In particular, the key idea is that highlighting the ability of AI to generate seemingly real content encourages news consumers to engage more with a trusted outlet. We partner with a large news outlet in Germany to test this by encouraging users to engage with custom-created content. In particular, as part of the field experiment, we utilize the mailing list of our partner news organization. We use approximately 100,000 subscribers as part of the experiment, who are then randomly split into three groups. In the treatment group, readers gain access to an AI literacy article that highlights ways to identify deepfake videos better. In this process, readers are encouraged to take a quiz that tests their ability to identify AI-generated videos and based on each question, the article will provide dimensions of videos that readers should look out for.
This intervention will test an active approach that can be taken by a news outlet to combat technology-enabled misinformation. The direction and magnitude of the results can have an important impact on how we design policies to combat misinformation and have individuals engage with high-quality news, which in turn provides incentives to news outlets to keep investing in writing stories with important societal implications.
As a benchmark to quantify the treatment effects, we will also have two control conditions. In the first control condition, we will email another group a link to a different, yet comparable, news story from SZ. This will account for changes in browsing and subscription behavior resulting from an email from the news organization itself. A final control group will receive no email, and we will track their browsing and subscription behavior as a "pure control" group.
|
After
This study aims to understand the impact of providing information about the potential of AI to generate misinformation (deepfakes) on news consumption and subscription outcomes. In particular, the key idea is that highlighting the ability of AI to generate seemingly real content encourages news consumers to engage more with a trusted outlet. We partner with a large news outlet in Germany to test this by encouraging users to engage with custom-created content. In particular, as part of the field experiment, we utilize the mailing list of our partner news organization. We use approximately 150,000 subscribers as part of the experiment, who are then randomly split into three groups. In the treatment group, readers gain access to an AI literacy article that highlights ways to identify deepfake videos better. In this process, readers are encouraged to take a quiz that tests their ability to identify AI-generated videos and based on each question, the article will provide dimensions of videos that readers should look out for.
This intervention will test an active approach that can be taken by a news outlet to combat technology-enabled misinformation. The direction and magnitude of the results can have an important impact on how we design policies to combat misinformation and have individuals engage with high-quality news, which in turn provides incentives to news outlets to keep investing in writing stories with important societal implications.
As a benchmark to quantify the treatment effects, we will also have two control conditions. In the first control condition, we will email another group a link to a different, yet comparable, news story from SZ. This will account for changes in browsing and subscription behavior resulting from an email from the news organization itself. A final control group will receive no email, and we will track their browsing and subscription behavior as a "pure control" group.
|
|
Field
Last Published
|
Before
September 29, 2025 10:57 AM
|
After
October 09, 2025 02:02 PM
|
|
Field
Intervention (Public)
|
Before
We use approximately 100,000 subscribers as part of the experiment, who are then randomly split into three groups. In the treatment group, readers gain access to an AI literacy article that highlights ways to identify deepfake videos better as they go through a quiz.
As a benchmark to quantify the treatment effects, we will also have two control conditions. In the first control condition, we will email a group a link to a different SZ news story. This will account for changes in browsing and subscription behavior resulting from an email from the news organization itself. A final control group will receive no email, and we will track their browsing and subscription behavior as a "pure control" group.
|
After
We use approximately 150,000 subscribers as part of the experiment, who are then randomly split into three groups. In the treatment group, readers gain access to an AI literacy article that highlights ways to identify deepfake videos better as they go through a quiz.
As a benchmark to quantify the treatment effects, we will also have two control conditions. In the first control condition, we will email a group a link to a different SZ news story. This will account for changes in browsing and subscription behavior resulting from an email from the news organization itself. A final control group will receive no email, and we will track their browsing and subscription behavior as a "pure control" group.
|
|
Field
Experimental Design (Public)
|
Before
We use approximately 100,000 subscribers as part of the experiment, who are then randomly split into three groups. In the treatment group, readers gain access to an AI literacy article that highlights ways to identify deepfake videos better. They will be guided through different techniques to spot deepfakes as they are encouraged to take a quiz that tests their ability to identify AI-generated videos.
We will also have two control conditions. In the first control condition, we will email a group of individuals on the mailing list a link to a different SZ news story. This will account for changes in browsing and subscription behavior resulting from an email from the news organization itself. A final control group will receive no email, and we will track their browsing and subscription behavior as a "pure control" group.
We will employ a between-subjects design, such that an individual on the mailing list will be assigned to one of the treatments. The main outcome variables of interest will be their subsequent browsing and subscription behavior.
We will also explore heterogeneity along two dimensions: (1) their prior news consumption and (2) their prior knowledge based on topics they consume on the platform
|
After
We use approximately 150,000 subscribers as part of the experiment, who are then randomly split into three groups. In the treatment group, readers gain access to an AI literacy article that highlights ways to identify deepfake videos better. They will be guided through different techniques to spot deepfakes as they are encouraged to take a quiz that tests their ability to identify AI-generated videos.
We will also have two control conditions. In the first control condition, we will email a group of individuals on the mailing list a link to a different SZ news story. This will account for changes in browsing and subscription behavior resulting from an email from the news organization itself. A final control group will receive no email, and we will track their browsing and subscription behavior as a "pure control" group.
We will employ a between-subjects design, such that an individual on the mailing list will be assigned to one of the treatments. The main outcome variables of interest will be their subsequent browsing and subscription behavior.
We will also explore heterogeneity along two dimensions: (1) their prior news consumption and (2) their prior knowledge based on topics they consume on the platform
|
|
Field
Planned Number of Observations
|
Before
About 100,000 -- this could be subject to some change based on survey uptake and traffic allocation by the company.
|
After
About 150,000 -- this could be subject to some change based on survey uptake and traffic allocation by the company.
|
|
Field
Sample size (or number of clusters) by treatment arms
|
Before
30,000
|
After
50,000
|