💰 quantitative finance

Auditing the Auditors: Does Community-based Moderation Get It Right?

This paper critiques the consensus-based auditing mechanism in X's Community Notes for inducing strategic conformity among minority contributors and proposes a novel two-stage algorithm that weights users by the stability of their past evaluations rather than agreement with the majority, thereby improving predictive accuracy while preserving dissenting voices.

Original authors: Yeganeh Alimohammadi, Karissa Huang, Christian Borgs, Jennifer Chayes

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yeganeh Alimohammadi, Karissa Huang, Christian Borgs, Jennifer Chayes

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, chaotic town square where thousands of people are trying to figure out which rumors are true and which are fake. To help, the town has set up a "Community Note" system: anyone can write a sticky note on a rumor to explain the truth, and everyone else can vote on whether that note is helpful.

The big question is: How do we decide who gets to keep writing notes, and whose votes count the most?

This paper investigates a specific rule used by a major social platform (X, formerly Twitter) and finds that the rule is actually making the town square less smart, not more. Here is the breakdown in simple terms.

1. The Problem: The "Popularity Contest" Trap

The platform introduced a rule called Consensus-Based Auditing. Think of it like a game where your score depends on how much you agree with the final group decision.

How it works: If you rate a note as "Helpful," and the group eventually decides that note was helpful, you get points. If you disagreed with the group, you lose points. If you lose too many points, you get kicked out of the game.
The Intention: The idea was to reward people who are "accurate" and punish those who are "wrong."
The Reality: It turned into a Popularity Contest.

2. The Consequence: The "Silent Minority"

The authors found that this rule created a strange behavior called Strategic Conformity.

Imagine you are a person in the town square who knows a rumor is false. But you look around and see that 60% of the crowd is about to vote that the rumor is true.

Under the old system: You would vote "False" because that's what you believe.
Under the new system: You think, "If I vote 'False,' I'll lose my points and get kicked out. If I vote 'True' (even though I know it's wrong), I keep my job."

So, you vote "True."

The Result:

The Minority Shrinks: People with minority opinions stop speaking up or start pretending to agree with the majority just to stay in the game.
The Controversy Gap: The system works fine for boring topics (like "The sky is blue"), but it fails miserably on controversial topics (like politics or war). On these topics, the "truth" is often split. Because the system punishes disagreement, fewer people write notes on controversial topics, and the notes that do appear are often just the majority view, ignoring important alternative perspectives.
The Crystal Ball Breaks: Because everyone is faking their votes to match the crowd, the computer algorithm trying to predict "what is true" gets confused. It starts making more mistakes because the data it's fed is no longer honest; it's just a reflection of what people think the crowd wants.

3. The Proposed Solution: The "Stable Judge"

The authors propose a new way to run the town square. Instead of asking, "Did you agree with the crowd?", they ask, "Are you a consistent thinker?"

They suggest a Two-Stage Algorithm:

Stage 1 (The Guess): The computer makes a quick guess about what the truth is based on everyone's votes.
Stage 2 (The Audit): The computer looks at how much each person's votes deviated from that guess.
- If Person A votes wildly differently every time (sometimes saying "Yes," sometimes "No" to the same thing), they are unstable and get less influence.
- If Person B votes consistently (even if they consistently disagree with the majority!), they are stable and get more influence.

The Metaphor:
Imagine a panel of judges at a cooking competition.

Current System: The judges are only allowed to keep their jobs if they vote for the dish the audience liked most. If they vote for a dish the audience hated, they are fired. Naturally, judges start voting for whatever the audience likes, even if they think the food is terrible. The competition becomes a popularity contest, not a taste test.
New System: The judges are judged on their consistency. If a judge always says, "This soup is too salty," and the soup is consistently too salty (even if the audience loves it), that judge is a good judge. If another judge says "This soup is great" one day and "This soup is terrible" the next, they are a bad judge, even if they agree with the audience sometimes.

Why This Matters

The paper argues that for a system to find the truth, it needs diverse, independent voices, especially on controversial issues.

If you only listen to the majority, you create an Echo Chamber.
If you reward people for being predictably honest (even when they are in the minority), you get a more accurate picture of reality.

In short: The current system rewards people for "fitting in." The new system rewards people for "thinking clearly." By switching to the new system, the platform could stop punishing minority viewpoints and actually get better at spotting misinformation.

1. Problem Statement

Online platforms increasingly rely on crowdsourced systems (e.g., X's Community Notes) to identify and label misleading content at scale. A critical design challenge is auditing the auditors: determining which user evaluations are reliable enough to influence the final platform outcome.

Current systems, specifically X's Community Notes since September 2022, employ Consensus-Based Auditing. In this design:

Users rate the helpfulness of notes.
A matrix factorization algorithm aggregates these ratings to determine a "consensus" (whether a note is displayed).
Auditing Rule: A user's future eligibility and influence are tied to their historical alignment with this eventual consensus. Users who agree with the final outcome gain influence; those who disagree are downweighted or lose eligibility.

The Core Hypothesis: The authors argue that this design creates a perverse incentive structure. By rewarding agreement with the eventual aggregate, the system encourages strategic conformity. Contributors, particularly those with minority viewpoints, may anticipate the platform's outcome and adjust their ratings to match the majority rather than providing independent, truthful signals. This leads to the suppression of diverse viewpoints and reduced coverage of controversial topics where misinformation is most prevalent.

2. Methodology

The paper employs a tripartite methodology: Empirical Analysis, Behavioral Modeling, and Algorithmic Proposal.

A. Empirical Analysis

Data: The authors utilize the publicly released Community Notes dataset (June 2022 – May 2023), focusing on the period surrounding the September 2022 rollout of "Rating Impact" (the consensus-based auditing rule).
Reconstruction: Since the platform does not release latent parameters, the authors re-implemented X's open-source matrix factorization algorithm to reconstruct weekly estimates of user factors ( $f_u$ ) and note factors ( $g_n$ ).
Metrics:
- Latent Factor Shifts: Analyzed changes in the distribution of user ideological factors before and after the policy change.
- Predictive Power: Measured the correlation between user-note alignment ( $f_u \cdot g_n$ ) and helpfulness ratings.
- Topic Engagement: Compared participation rates on "controversial" vs. "non-controversial" topics using both topic-based (LLM/Bag-of-Words) and factor-based classifications.
- Model Performance: Evaluated out-of-sample prediction error (MSE) of the platform's latent-factor model.

B. Behavioral Modeling

The authors developed a game-theoretic model to explain the empirical shifts:

Utility Function: Users maximize a utility function balancing their private signal ( $r^*_{un}$ ) against a penalty for deviating from the anticipated consensus ( $\tilde{m}_n$ ).
Strategic Reporting: The optimal report $a^*_{un}$ is a convex combination:
$a^*_{un} = \rho(c_n)r^*_{un} + (1 - \rho(c_n))\tilde{m}_n(c_n)$
Where $\rho(c_n)$ is a conformity weight that decreases as the controversy ( $c_n$ ) of the note increases.
Theoretical Implication: The model proves that when $\rho < 1$ , the matrix factorization algorithm recovers a conformity-distorted projection of the truth rather than the true latent helpfulness. Specifically, minority user factors are systematically biased toward the center or the majority, leading to a "compression" of the minority share.

C. Proposed Solution: Two-Stage Weighted Matrix Factorization

To mitigate these issues, the authors propose an alternative auditing algorithm that decouples influence from consensus agreement:

Stage 1: Fit a standard regularized matrix factorization model to estimate latent factors and compute residuals ( $e_{un}$ ) for each user-note pair.
Stage 2: Estimate the variance of residuals ( $\hat{\sigma}^2_u$ ) for each user. This measures the stability and predictability of a user's evaluations relative to the latent structure, not their agreement with the majority.
Weighted Aggregation: Refit the matrix factorization model using inverse-variance weights ( $w_u = 1/\hat{\sigma}^2_u$ $w_{u} = 1/ \overset{σ}{^}_{u}^{2}$ ).
- Users with stable, informative residuals (even if they consistently disagree with the majority) retain high influence.
- Users with erratic or noisy evaluations are downweighted.

3. Key Results

Empirical Findings

Strategic Conformity: Post-rollout, the distribution of user factors shifted. The share of "minority-aligned" raters decreased from 41.6% to 36.2%, and their factors moved closer to the center (mean absolute factor dropped from 0.522 to 0.413).
Reduced Predictive Power: The correlation between user-note alignment and helpfulness ratings dropped significantly (Spearman correlation fell from 0.792 to 0.525), indicating that alignment became a weaker predictor of truth.
Suppression of Controversial Topics:
- Notes on controversial topics (politics, conflict) saw a relative decline in "Helpful" status compared to non-controversial topics.
- Minority users specifically reduced their engagement with controversial content by ~7.5 percentage points relative to non-minority users post-rollout.
Model Degradation: The platform's latent-factor model suffered a 116% increase in in-sample MSE and a 76% increase in one-week-ahead MSE after the policy change, suggesting the data became noisier and less predictable due to strategic behavior.

Algorithmic Performance

The proposed Two-Stage Weighted Matrix Factorization significantly outperformed the baseline.
It reduced the Mean Absolute Residual by 5.73% and the Median Absolute Residual by 27.99% in out-of-sample (one-week-ahead) predictions.
Crucially, this method preserved the influence of consistent minority contributors who disagreed with the consensus, whereas the baseline method penalized them.

Theoretical Proofs

Theorem 0.2: Proves that under consensus-based auditing ( $\rho < 1$ ), the estimator for note helpfulness is inconsistent and biased ( $\hat{i}_n \to i^\infty_n \neq i_n$ ).
Proposition 0.5: Demonstrates that the estimated share of the minority shrinks ( $\pi_{est} < \pi_{true}$ ) as conformity increases.
Theorem 0.7: Proves that the two-stage inverse-variance weighting yields a consistent estimator with the minimum asymptotic variance among weighted matrix factorization estimators, provided users report private signals.

4. Key Contributions

Empirical Evidence of Strategic Conformity: Provided the first large-scale empirical quantification showing that consensus-based auditing rules in crowdsourced moderation systematically distort contributor behavior, causing minority voices to self-censor or align artificially.
Behavioral Mechanism: Formalized a behavioral model showing that tying eligibility to consensus creates a "Spiral of Silence" effect, where the platform's own aggregation logic becomes a self-fulfilling prophecy that distorts the truth.
Novel Auditing Algorithm: Designed and validated a residual-stability-based auditing mechanism. This approach shifts the metric of reliability from "agreement with the majority" to "statistical consistency of private signals," theoretically allowing for accurate aggregation even in the presence of diverse, dissenting viewpoints.

5. Significance

This paper challenges a fundamental assumption in content moderation: that aggregating user ratings based on consensus is the optimal way to identify truth.

Design Implication: It suggests that auditing rules are not just administrative details but core system components that shape the information ecosystem. Rewarding agreement can destroy the "wisdom of crowds" by eliminating the independence required for that wisdom to function.
Practical Application: The proposed two-stage algorithm offers a concrete, statistically grounded alternative for platforms (like X, Meta, TikTok) to improve the accuracy of fact-checking systems, particularly on polarized topics where misinformation is most dangerous.
Broader Impact: The findings highlight the tension between operational efficiency (using consensus as a proxy for truth) and epistemic diversity (needing disagreement to find errors). The paper argues that preserving independent, minority viewpoints is essential for effective large-scale moderation.