An experimental study of the influence of anonymous information on social media users

TL;DR

This study asks whether anonymous information can sway social-media-like opinion choices. Using a two-phase preregistered online experiment with inkblot-based stimuli and four information conditions, the authors quantify how anonymous comments alter second-phase opinions and show that such influence can affect up to about half the participants, especially when initial confidence is low. They complement the empirical findings with Bayesian multilevel multinomial regression and agent-based modeling to reveal that simple mechanisms can reproduce the observed shifts toward promoted opinions, and that confidence attenuates this effect. The work highlights that even ostensibly anonymous feedback can reshape opinion distributions and ranking, carrying implications for online information ecosystems and moderation strategies.

Abstract

Increasingly, people use social media for their day-to-day interactions and as a source of information, even though much of this information is practically anonymous. This raises the question: does anonymous information influence its recipients? We conducted an online, two-phase, preregistered experiment using a nationally representative sample of participants from the U.S. to find the answer. To avoid biases of opinions among participants, in the first phase, each participant examines ten Rorschach inkblots and chooses one of four opinions assigned to each inkblot. In the second phase, the participants are randomly assigned to one of four distinct information conditions and are asked to revisit their opinions for the same ten inkblots. Conditions ranged from repeating phase one to receiving anonymous comments about certain opinions. Results were consistent with the preregistration. Importantly, anonymous comments shown in phase two influence up to half of the participants' opinion selections. To better understand the role of anonymous comments in influencing the selections of opinions, we implemented agent-based modeling (ABM). ABM results suggest that a straightforward mechanism can explain the impact of such information. Overall, our results indicate that even anonymous information can have a significant impact on its recipients, potentially altering their popularity rankings. However, the strength of such influence weakens when recipients' confidence in their selections increases. Additionally, we found that participants' confidence in the first phase is inversely related to the number of change opinions.

Figures (13)

• Figure S1: Inkblot examples at different entropies. This figure shows three of the inkblots presented in our surveys. For each inkblot we give the entropy for the first phase opinion counts, the average confidence of the inkblot, and the list of opinions the participants could choose from. The average confidence is represented numerically by encoding the confidence categories into values 1 - 4 with 1 being the least and 4 being the most confident. For each opinion, we provide the fraction of participants who chose the opinion as well as their average confidence. The opinions inherit their rank from the place of their popularity on the list of all opinions' popularity's ordered from the highest to the lowest.
• Figure S2: Predicted opinion transitioning frequency by condition. This figure shows the results focusing on the relationship between condition, opinion, and predicted transitioning frequency. In this figure we can see that each influencing condition has some significant effect on reducing the amount of transitioning that occurs for participants holding the promoted opinion of the condition in the first phase.
• Figure S3: Categorical Regression Model, all terms
• Figure S4: Opinion transition matrix. This figure defines the fractional flow of opinions, $O_i^1 \to O_j^2$ for some opinion $i$ in the first phase to some opinion $j$ in the second phase, for each condition.
• Figure S5: Opinion transition matrices for CR, CM, and CS conditions compared with C. This figure defines the fractional flow of opinions, $O_i^1 \to O_j^2$ for some opinion i in the first phase to some opinion j in the second phase, relative to such frequencies from the C condition. A positive value indicates a greater frequency versus control while a negative value indicates the opposite. The highlighted column in each subplot (bounded in yellow) refers to the promoted opinion of that condition.