When U.S. Liberals Demand More Censorship

Please see Hidden.

2024-04-08

2024-04-26

Demand for censorship on social media content that does not contain misinformation

Each of the 10 outcome posts has two versions, one with misinformation and one without misinformation. Each participant will only see one of the two versions for each post, randomly selected. For each post, participants will be asked to rate their censorship preference on a 1~5 scale, with 1 denoting definitely not censoring and 5 denoting definitely censoring. For each participant, the primary outcome (demand for censorship on social media content that does not contain misinformation) will be calculated as the mean of their responses to the censorship preference questions on non-misinformation outcome posts. We will also measure this as a binary variable by examining whether this mean is greater than 4, indicating a high preference for censorship of non-misinformation posts.

1) Demand for censorship on social media content that contains misinformation; 2) Demand for censorship on social media in general; 3) Perception of negative externalities of information; 4) Combined outcomes of perceived negative externalities and demand for censorship

For each participant, the first outcome (demand for censorship on social media content that contains misinformation) will be calculated as the mean of their responses to the censorship preference questions on misinformation outcome posts, and the second outcome (demand for censorship on social media in general) will be calculated as the mean of their responses to the censorship preference questions on all outcome posts. We will also measure these as binary variables by examining whether each mean is greater than 4.

A random half of the participants will see one additional question for each post, on the extent to which they find the post may negatively affect other users on the social media platform. For each participant, the third outcome (perception of negative externalities of information) will be based on the mean of their responses to the negative effect questions for non-misinformation posts and misinformation posts respectively, binarized based on whether the mean is greater than 4. The fourth outcome (combined outcomes of perceived negative externalities and demand for censorship) will be the mean of the products of each participant's corresponding responses of demand for censorship and binarized perceived negative effects.

Please see Hidden.

Participants will be included in the survey experiment if they indicate their ideology to be liberal (1 or 2 on a scale of 1~5). Then, admitted participants will be told that they are participating in a survey run by a social media company seeking to base its content moderation algorithm on their responses. Each participant will see 20 social media posts, the first 10 differing between treatment and control groups, and the next 10 as outcomes. For the control group, the first 10 posts will concern a wide array of political topics framed from liberal/left-wing perspectives. For the treatment group, the first 10 posts will feature the same topics, but with content that emphasizes the negative externalities resulting from conservative/right-wing views on these issues. The next 10 posts reflect conservative/right-wing perspectives of these issues, and are used to measure demand for social media censorship. Each of the 10 outcome posts has two versions, one with misinformation and one without misinformation. Each participant will only see one of the two versions for each post, randomly selected. For each post, participants will be asked to rate their censorship preference on a 1~5 scale, with 1 denoting definitely not censoring and 5 denoting definitely censoring. A random half of the participants will see one additional question for each post, on the extent to which they find the post may negatively affect other users on the social media platform.

Randomization will be conducted by Qualtrics.

individual

No

N/A

We will recruit 2,000 participants (Americans who self-identify as liberals ideologically) via Lucid.

Half of the 2,000 participants will be randomly assigned to treatment and control respectively, leading to approximately 1,000 in each group.

For continuous measures, we will be able to detect an effect size of 0.15 at 90% confidence level with 80% power, with a control baseline of 2.5 (on a scale of 1~5) and standard deviation 1.3. For discrete measures, we will be able to detect an effect size of 3.5 percentage points at 90% confidence level with 80% power, with a control baseline of 10%. In our regression analysis, we will also control for the following pre-treatment covariates: demographics (age, sex, LGBT status, race, college education, religion), political trust, media consumption, and social media uses and habits. We expect controlling for these covariates to increase our power.