Algorithm-Driven Information Similarity and Collective Action: An Experimental Study

Information Similarity Reverses Collective Action: Coordination vs. Free-Riding Depends on Goal Difficulty

When people draw on a common source of information, such as social media, they become more confident about what others have seen. This shared awareness can help a group coordinate on a common goal. However, it can also tempt individuals to "free-ride" (letting others shoulder the cost of action while reaping the benefits).

A new study from Peking University explores this tension in the context of digital collective action. The researchers investigate how the similarity of information across a group shapes whether people cooperate or stay idle. They find that the impact of shared information is not universal. Instead, it flips based on how difficult the collective goal is to achieve.

The Failure of Uniform Information Models

Current understanding of digital mobilization often treats information flow as a monolithic force. Social media algorithms push similar content to millions of users. This can ignite massive protests or dampen individual participation. Existing models struggle to explain why the same algorithmic alignment leads to such contradictory outcomes.

Previous studies have provided conflicting evidence. Some suggest common information sources mobilize citizens. Others find that learning about others' intentions makes individuals less likely to act. The missing link is the institutional "conversion channel" (the rules determining how individual efforts combine into a successful outcome). Without accounting for the difficulty of the goal, we cannot predict if shared information fosters cooperation or encourages laziness.

Testing the Strategic Channel of Similarity

To isolate this effect, the authors designed a content-moderation game involving 576 participants. The setup mimics a real-world platform. Users decide whether to pay a personal cost to report harmful content. The content is only removed if the number of reports meets a specific threshold ($T$).

The researchers manipulated information similarity using a parameter $\rho$ (rho). This represents the degree of correlation between reports sent to different group members. The mechanism works in two distinct branches: 1. Common Evaluation: With probability $\rho$, the algorithm sends the identical report to all twelve members. 2. Independent Evaluation: With probability $1-\rho$, the algorithm conducts twelve separate, independent evaluations.

Crucially, the authors held the "marginal informativeness" (the accuracy of a single person's report) fixed. This allows the study to isolate the strategic effect—how much a person learns about others—from the fundamental effect of what they learn about the world. As shown in the experimental protocol, subjects receive their report and answer belief questions before making their decision.

A Reversal in Participation and Welfare

The paper reports a striking sign reversal in how similarity affects behavior. The effect of information similarity on reporting depends entirely on the removal threshold. The authors find that increasing similarity lowers reporting by 17 percentage points when the threshold is low (an "easy" goal). Conversely, it raises reporting by 34 percentage points when the threshold is high (a "demanding" goal) .

This reversal is driven by "pivotality" (the probability that a single individual's action will be the one to tip the scales). The authors measure this through elicited beliefs. They find that: * In Low-threshold regimes, similarity leads to "unrecognized pivotality." People feel less influential and report less, even though their potential to be pivotal has actually increased. * In High-threshold regimes, similarity leads to "illusory pivotality." People feel more influential and report more, even though their actual chance of being pivotal is nearly zero.

This mismatch between perception and reality has serious consequences for social welfare. The authors report that welfare (the total benefit of removal minus the cost of reporting) falls under both miscalibrated patterns. In easy regimes, welfare drops because people free-ride and miss reachable goals. In demanding regimes, welfare drops because people spend resources on "wasted" reports that fail to clear the high bar .

Limitations of the Laboratory Proxy

While the experimental design is rigorous, it possesses inherent boundaries. The study is conducted in a controlled laboratory environment. It may not fully capture the complex dynamics of real-world social media ecosystems. In those settings, information similarity is often bundled with changes in content sentiment or network structure.

Furthermore, the authors note that the "high-threshold" effect might be a boundary condition. This effect occurs where mobilization fails to translate into success. This might be specific to finite group sizes. In a massive, global network, the relationship between individual pivotality and aggregate success might behave differently than in a group of twelve.

The Verdict: Design Matters More Than Data

Is information similarity a tool for coordination or a catalyst for free-riding? According to this study, the answer depends on your rules.

If you are designing a system for collective action, you cannot look at the information architecture in isolation. If your threshold for success is too low, algorithmic alignment will inadvertently encourage free-riding. If your threshold is too high, it will trigger a surge of "wasted" effort. This effort consumes resources without achieving the goal. To maximize social welfare, practitioners must ensure the institutional threshold is "aligned" with the information structure. This creates a regime where perceived influence accurately reflects actual opportunities to make a difference.

Figures from the paper