💻 computer science

Hyperactive Minority Alters the Stability of Community Notes

This paper analyzes X's Community Notes system and finds that its stability is compromised by a small, highly polarized minority of active users whose disproportionate influence determines which notes appear, effectively reconfiguring rather than decentralizing epistemic authority.

Original authors: Jacopo Nudo, Eugenio Nerio Nemmi, Edoardo Loru, Alessandro Mei, Walter Quattrociocchi, Matteo Cinelli

Published 2026-02-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jacopo Nudo, Eugenio Nerio Nemmi, Edoardo Loru, Alessandro Mei, Walter Quattrociocchi, Matteo Cinelli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive town square where everyone is shouting. In this square, there's a special rule: if someone says something that might be a lie, anyone can write a sticky note explaining why it's false. Then, other people in the square vote on whether that sticky note is helpful. If enough people agree it's helpful, the note gets plastered right next to the original shout for everyone to see.

This is Community Notes on X (formerly Twitter). The idea sounds beautiful and democratic: "The people will fact-check the people!"

But a new study by researchers in Rome reveals a surprising twist. The system isn't run by the "wisdom of the crowd." Instead, it's being run by a tiny, hyper-active, and very opinionated minority.

Here is the breakdown of what they found, using some simple analogies:

1. The "1% Rule" on Steroids

In most social media, a small group of people does most of the posting. But in Community Notes, the imbalance is extreme.

The Analogy: Imagine a classroom of 100 students. You'd expect maybe 20 kids to raise their hands to vote on a class rule. But in this study, they found that 20 kids are doing 80% of the voting.
The Reality: A tiny fraction of users (less than 1%) are responsible for the vast majority of the ratings. The system isn't a broad consensus; it's a loud echo chamber created by a few dozen people who never stop clicking "Helpful" or "Not Helpful."

2. The "Super-Partisan" Voters

These super-active voters aren't neutral referees. They are deeply polarized.

The Analogy: Think of a sports game. You want the referees to be neutral. But in this study, the referees who blow the whistle the most are actually die-hard fans of one specific team. They only watch games involving their team, and they only vote against the other team.
The Reality: The people who rate the notes the most are significantly more politically extreme than the average user. They tend to rate notes that attack their political opponents as "Helpful" and notes that attack their own side as "Not Helpful." They aren't looking for truth; they are looking for ammunition.

3. The House of Cards

This is the most dangerous part. The researchers ran a simulation where they pretended these super-active voters didn't exist. They asked: "What if we just removed the top 10, 100, or 1,000 most active voters?"

The Analogy: Imagine a Jenga tower. You think it's a solid structure built by thousands of blocks. But when you pull out just the top 10 blocks (the ones holding everything up), the whole tower collapses or rearranges itself completely.
The Reality: When they removed just a tiny fraction of these hyper-active users, the list of "Helpful" notes changed drastically. Notes that were visible to the whole world suddenly disappeared, and new notes appeared. The system is so unstable that the entire "truth" of what gets shown depends on the mood of a few dozen people on a given day.

The Big Picture

The paper concludes that Community Notes was supposed to be a democracy, but it has accidentally become an oligarchy (rule by a few).

Instead of decentralizing power (spreading it out to everyone), the system has concentrated power in the hands of a small, polarized group. They act as the gatekeepers of truth, but because they are so active and so biased, they can easily skew the system to show only the "facts" that fit their political narrative.

In short: The "wisdom of the crowd" is a myth here. It's actually the "whim of a few" that decides what you see.

1. Problem Statement

The paper addresses the growing reliance on community-based fact-checking systems (specifically Community Notes on X, formerly Twitter) as a scalable alternative to professional fact-checking. While these systems are designed to be transparent, democratic, and consensus-driven, the authors question whether the underlying social dynamics of the user base undermine the system's stability and neutrality.

The core research problem is to determine:

To what extent are rating activities concentrated among a small minority of users?
Are these highly active users politically neutral, or do they exhibit polarization?
How sensitive is the consensus algorithm to the participation of this "hyperactive" minority? Does the system's output (which notes become visible) depend disproportionately on a few dozen users?

2. Methodology

The authors employed a large-scale empirical analysis combined with counterfactual simulations using the following data and tools:

Dataset: The complete public release of Community Notes data from January 28, 2021, to December 2, 2025.
- Scale: ~2.2 million notes, ~1.3 million distinct contributors/raters, and ~184 million rating events.
- Political Subset: For polarization analysis, they utilized a subset of ~218,000 notes linked to tweets from accounts with inferred partisan labels (Republican or Democrat), covering ~18.2 million rating events.
Algorithm Replication: The authors integrated the open-source implementation of the official Community Notes consensus algorithm used in production. This allowed them to replicate the system's logic and run simulations.
Analytical Techniques:
- Power-Law Modeling: Used to characterize the distribution of rating activity (fitting a discrete power-law distribution $P(X=x) \propto x^{-\alpha}$ ).
- Inequality Metrics: Calculated the Gini coefficient to quantify the inequality of rating distribution.
- Selectivity Analysis: Modeled user behavior using a saturation function $N(r) = N_{asy}(1 - e^{-r/\tau})$ to measure how quickly raters focus on a limited set of authors compared to a randomized baseline.
- Political Leaning Inference: Defined a political leaning score ( $L \in [-1, 1]$ ) based on the interaction between the tweet author's party, the note's classification (Misleading vs. Not Misleading), and the rater's vote (Helpful vs. Not Helpful).
- Counterfactual Simulations: Systematically removed the top $N$ most active raters (from 10 to 10,000 users) from the input dataset and re-ran the consensus algorithm to observe changes in note status.
- Stability Metric: Used Jaccard Similarity to measure the overlap between the set of visible notes in the baseline scenario versus the simulation scenarios.

3. Key Contributions

Empirical Characterization of Activity Concentration: The study provides the first large-scale evidence that Community Notes activity follows a heavy-tailed distribution, where a tiny minority of users generates the vast majority of ratings.
Link Between Activity and Polarization: It demonstrates that the most active contributors are not a representative sample of the community; they are significantly more politically polarized and selective in the content they engage with compared to the general population.
Algorithmic Sensitivity Analysis: By leveraging the open-source algorithm, the authors prove that the system is structurally unstable. The visibility of notes is highly sensitive to the participation of a small, hyperactive group, meaning the "consensus" is often driven by a few dozen individuals rather than a broad cross-section of the community.

4. Key Results

A. Extreme Concentration of Activity

Pareto Principle: The distribution of ratings follows a power law with a scaling exponent $\alpha \approx 3.06$ .
80/20 Rule: Approximately 20% of users account for 80% of all ratings.
Gini Coefficient: The inequality in rating activity is high ( $G = 0.78$ ) and has increased over time, stabilizing at a high level after 2022.
Selectivity: Highly active raters do not rate broadly; they concentrate their activity on a small subset of tweet authors. The characteristic scale ( $\tau$ ) for reaching new authors is significantly lower for real users than for randomized baselines, indicating a strong preference for specific targets.

B. Political Polarization of the Hyperactive Minority

Polarization Gradient: As user activity increases, political polarization intensifies.
Decile Analysis: Users in the top decile of activity exhibit a strong Republican-leaning bias (positive $L$ values), whereas the overall population is relatively balanced (slightly Democratic-leaning).
Bimodality: The distribution of political leaning among top contributors is highly bimodal (Ashman's $D > 3.6$ ), indicating distinct, opposing camps, whereas less active users show more mixed ratings.

C. Systemic Instability (Counterfactual Results)

Sensitivity to Removal: Removing a minuscule fraction of users drastically alters the system's output.
- Removing the top 10 most active users caused 4,587 notes previously marked as "Currently Rated Helpful" (CRH - publicly visible) to lose that status.
- Removing 10,000 users (approx. 0.007% of the rater pool) reduced the Jaccard similarity of "Not Helpful" (CRNH) statuses to 69.6%.
Conclusion: The emergence and visibility of notes are not robust; they depend heavily on the behavior of a few dozen highly active, polarized users.

5. Significance and Implications

Reconfiguration of Epistemic Authority: The paper argues that community-based fact-checking does not decentralize authority as intended. Instead, it reconfigures it, concentrating power in the hands of a small, polarized, and hyperactive minority.
Threat to Neutrality: Because the most influential raters are politically biased and selective, the system risks amplifying partisan dynamics rather than neutralizing misinformation. The "consensus" may reflect the views of a specific faction rather than a broad cross-partisan agreement.
Structural Vulnerability: The system is fragile. Minor perturbations in the participation of the top contributors can lead to markedly different outcomes regarding which corrections are visible to the public.
Policy Recommendations: The authors suggest that for community-based moderation to be effective, it must address participation inequalities. Potential solutions include capping individual activity levels, dynamically down-weighting hyperactive users, or re-evaluating the consensus algorithm to ensure robustness against participation asymmetries.

In summary, the paper challenges the narrative that Community Notes represents a "wisdom of the crowd" solution, revealing instead a system where the "crowd" is effectively a small, polarized elite whose behavior dictates the visibility of truth on the platform.