Algorithm-Driven Information Similarity and Collective Action: A Laboratory Experiment
How to cite
Xia, Bozhang; Khanna, Manshu. 2026. "Algorithm-Driven Information Similarity and Collective Action: A Laboratory Experiment." AEA RCT Registry. AEARCTR-0018776. https://doi.org/10.1257/rct.18776-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
Algorithmic content curation increasingly shapes the information that individuals observe in online environments. Existing concerns about such systems often emphasize echo chambers and reduced exposure to diverse viewpoints. However, information similarity may also affect strategic behavior: when individuals receive more similar information, they may better anticipate what others know and how others will act. Whether this facilitates or undermines collective action depends on the nature of the underlying coordination problem.
This study examines how algorithm-driven information similarity affects collective action in a laboratory experiment. Participants make costly reporting decisions in a content-moderation game where successful removal of harmful content depends on collective participation. The experiment varies the similarity of algorithmic signals while holding individual signal accuracy fixed. It also varies the institutional threshold for successful collective action, creating environments in which others’ participation may either encourage one’s own participation or reduce the need to act. The study measures reporting decisions and beliefs about others’ behavior to examine when information similarity promotes coordination and when it exacerbates free-riding.
Intervention
The study is a laboratory experiment.
In our experiment, participants take part in a content-moderation game. In each round, participants are placed in a group and decide whether to report a piece of online content. Reporting is costly, while successful removal of harmful content benefits all group members. Content is removed only if the number of reports reaches a threshold. Before making their decisions, participants receive an algorithmic signal about whether the content is likely to be harmful. The experiment varies both the similarity of these algorithmic signals across group members and the threshold regime that determines how many reports are needed for removal.
We run the following threshold-regime treatments between sessions:
1. Low-threshold regime: Content can be removed with a relatively small number of reports. This environment is designed to capture situations in which collective action is relatively easy. This environment is designed to capture situations in which collective action is relatively easy and free-riding incentives may be more salient.
2. Medium-threshold regime: Content removal requires an intermediate number of reports.
3. High-threshold regime: Content removal requires a relatively large number of reports. This environment is designed to capture situations in which successful collective action requires broad coordination.
Within each session, the degree of algorithmic information similarity varies across rounds. Higher similarity means that group members are more likely to receive the same algorithmic signal, while lower similarity means that signals are more likely to be independently generated. Individual signal accuracy is held fixed, so variation in similarity changes participants’ uncertainty about others’ information rather than the informativeness of their own signal.
Participants also report incentivized beliefs about others’ reporting behavior before making their reporting decision. At the end of the experiment, participants complete a short survey and incentivized preference tasks.
Experimental design
The experiment employs a mixed laboratory design investigating how algorithmic information similarity affects collective action. Participants take part in a repeated content-moderation game in which they decide whether to report potentially harmful online content. Reporting is costly, and content is removed only if enough group members report it. Treatments vary the threshold regime for successful removal across sessions. Within sessions, the similarity of algorithmic signals varies across rounds while individual signal accuracy is held fixed. This design allows us to compare how information similarity affects reporting behavior across environments with different degrees of collective-action difficulty.
The experiment is designed to test the main predictions of Basak, Deb, and Kuvalekar (2026), "Similarity of Information and Collective Action" (American Economic Review): increasing information similarity should facilitate reporting when successful collective action requires broad coordination, but may reduce reporting when the threshold for success is low and free-riding incentives are more salient. The medium-threshold treatment is included to study the transition between these two cases.
Design details
Cited by
No citation data available for this registration yet.