Community Notes undermoderate polarizing content by design creating risks in electoral processes

Abstract

Community Notes of X enables users to collaboratively moderate misleading content. To resolve conflicting moderation, Community Notes infers a latent ideological dimension and selects notes garnering cross-partisan support. As this system is now deployed worldwide, we evaluate its operation across diverse polarization contexts. We analyze all 1.9 million moderation notes receiving 135 million ratings by March 2025, cross-referencing ideological scaling data on 13 countries. Our results show that the Community Notes algorithm effectively captures the main polarizing dimensions across countries, surfacing notes that garner cross-partisan support. This also means that, by design, Community Notes systematically undermoderate polarizing content. We analyze notes relating to four recent elections in the United States (2024), the United Kingdom (2024), France (2024), and Germany (2025) and demonstrate that they are systematically undermoderated when compared to other notes, posing potential risks to civic discourse and electoral processes.

New crowd-sourced moderation systems rolled out by platforms fail by design to address polarizing content.

INTRODUCTION

Social media platforms have become ubiquitous sources for the consumption of news content (1), where users may also participate in content creation and dissemination. A persistent challenge in these platforms is the dissemination of misinformation (2–4), prompting platforms to deploy means for content moderation (5, 6). While expert fact-checking has been shown to successfully decrease the extent to which people agree with misleading claims (7–9), this approach faces substantial limitations in practice: (i) professional fact-checking operations cannot scale to meet the volume of content requiring verification (10, 11), (ii) their efforts may be disconnected from online attention (12), (iii) moderation may be perceived as having partisan bias by the public (1), and (iv) they may suffer from low visibility (13). Additionally, professional moderation may backfire by further “entrenching” polarized individuals (14). On the other hand, research has shown that crowd-sourcing fact-checking assessments from regular users can effectively identify misinformation, with politically balanced laypeople ratings correlating strongly with those of professional fact-checkers (15–18).

Motivated by these findings, X (formerly Twitter) piloted a crowd-sourced fact-checking system in the United States in early 2021 (19), later deployed worldwide in 2023 under the name Community Notes (20). This system enables self-selected contributors to attach contextual notes to posts they consider “misinformed or potentially misleading.” Other contributors can then rate the helpfulness of these notes based on their subjective assessment of the content and its context. A core challenge for the platform is moderating polarizing content, for which consensus on usefulness may not emerge between raters across the ideological spectrum (e.g., when raters preferentially seek to moderate posts by political adversaries).

To address this challenge, X first learns to predict how contributors rate notes fitting latent ideology parameters for users and notes and then algorithmically identifies notes that appeal to users across ideological divides rather than only to like-minded individuals (19, 21). Instead of merely reflecting the idiosyncratic “helpfulness” of a note, contributor ratings are subjective assessments that may incorporate ideological alignment or misalignment between the rater and the claim being discussed. To account for this, X’s Community Notes system infers latent ideological positions for both the notes (θ_n) and the raters (θ_r) under the key hypothesis that greater ideological alignment increases the probability of a rater rating the note as Helpful.

Computationally, X first converts ratings given to notes to numerical values: 1 for Helpful, 0.5 for Somewhat Helpful, and 0 for Not Helpful (22). Then, to predict these empirical ratings, the system infers several parameters: note latent ideologies (θ_n), rater latent ideologies (θ_r), individual rater leniency (β_r), the tendency of notes to be rated helpful regardless of raters’ ideologies (β_n), and a global helpfulness baseline (β₀). These parameters are estimated, from the ratings alone, to best predict the rating η that rater r will assign to note n through the predictive formula ${\hat{\mathrm{\eta }}}_{\mathit{rn}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_n+{\mathrm{\beta }}_r+{\mathrm{\theta }}_n·{\mathrm{\theta }}_r$.

Consequently, a note rated Helpful by diverse raters (i.e., with θ_r values spread across positive and negative values) will have a higher β_n value and lower absolute value $\mid {\mathrm{\theta }}_n\mid$ when compared to notes rated Helpful by only one side of the spectrum. X’s Community Notes service uses these parameters to algorithmically resolve conflicting ratings by selecting notes with the highest β_n values (19) and displaying them under the post, visible to all users on the platform (23).

Results from X’s pilot phase in the United States revealed that notes selected for display garnered broad appeal across ideological positions (19) and that users exposed to such notes were less likely to share the posts and less inclined to agree with the posts’ misleading claims. Independent research found high agreement between these algorithmically selected notes and expert fact-checker classifications (24). Additionally, Allen et al. (25) showed that political partisanship was highly predictive of contributors’ ratings in the United States, with raters preferentially challenging content from those with whom they disagree politically. Following X’s Community Notes global deployment in 2023, subsequent studies corroborated these findings, documenting reduced sharing of flagged content (26) and accurate sourcing of credible references for notes addressing COVID-19 vaccination misinformation (4). Moreover, moderation resulting from the Community Notes service was shown to be less prone to backfire effects that entrench individuals (14, 27) and generated greater user trust than warnings issued by expert fact-checkers (28).

X’s Community Notes system was developed and tested in the United States (19, 25), where online polarization is traditionally modeled using a single ideology dimension (29, 30). However, the relevance of this single-dimensional representation underlying of notes and rating, central to the algorithmic moderation process in Community Notes, remains largely unconfirmed and unexplored in other national settings that might be structured along different ideology dimensions (31–33). Whether the Community Notes model captures relevant political dimensions in different countries has not been explored in previous works and has a direct impact in how the service functions as it determines what type of polarizing content will not be algorithmically moderated due to lack of sufficient agreement among raters. This study examines Community Notes usage on X and its algorithmic outcomes across all countries in which the service is available. We further assess the extent to which the Community Notes model captures ideological positioning in diverse national contexts and how this shapes downstream moderation outcomes not only in the United States, for which the system was conceived and tests, but also in a diversity of settings worldwide. Leveraging ideology scaling of follower networks (34) calibrated with political survey data (35), we compare the latent ideological parameter inferred by Community Notes algorithm with the main dimensions of political competition in 13 countries. Our results show that, in all 13 countries that we study, the Community Notes model reliably identifies the main ideological divide that structures political communication networks on X. This remains true even after accounting for a second dimension related to misinformation: trust in elites and institutions (36, 37). This implies that notes are proposed, rated, and surfaced as Helpful across the ideological spectrum. However, because this system only surfaces notes that receive broad cross-ideological agreement, content that strongly activates polarization is systematically undermoderated. We demonstrate the importance of this intrinsic limitation by showing that the share of notes related to elections is consistently lower than for other topics during four major elections in our study period: the 2024 US presidential election, the 2024 UK general election, the 2024 French legislative elections, and the 2025 German federal elections. While Community Notes offers several important benefits, this limitation poses a risk for information integrity during electoral processes.

Data and definitions

Community Notes data

To enable third-party scrutiny, X publicly discloses all proposed Community Notes, their ratings, the helpfulness status algorithmically assigned to them, and user requests soliciting Community Notes on posts that they believe would benefit from additional context (38).

As of 1 March 2025, X had disclosed data on 1.88 million Community Notes, which received 135.17 million ratings from 1.15 million contributors. Additionally, 1.49 million X users submitted 4.83 million requests for Community Notes associated with 1.91 million posts. Following X’s preprocessing protocol (19, 22), we retained notes with at least five ratings and contributors who rated at least 10 notes. This resulted in 1.45 million notes associated with 886 370 X posts, rated 133.20 million times by 821 250 contributors (see descriptive statistics in section S2.1); 59.5% of ratings classify the proposed notes as Helpful, 37.5% as Not Helpful, and 3.0% as Somewhat Helpful. To enable contextual analysis beyond ratings, the posts associated with 83.5% of proposed notes were retrieved, while the remainder had been deleted by March 2025 (see section S2.2). Hereafter, we use “notes” to refer to Community Notes flagging the original post as misinformed or potentially misleading (73.0% of all notes; the others are not meant for public display but allow contributors to argue why a note is not needed; see section S1).

Inferring the parameters of the Community Notes model

We applied the regularized matrix factorization outlined in X’s publication (19) and implemented in their open-source code (22), to infer the global helpfulness bias β₀, rater biases β_r, note biases β_n, and the latent ideologies of notes and raters, θ_n and θ_r (see details in section S6). The parameters mirror those reported in X’s publication (19) and align closely with note outcomes publicly displayed on X. Quantitatively, the note biases β_n we inferred from ratings predict whether a note was algorithmically selected as Helpful by X (hereinafter, achieving Helpful Status) with an AUC-ROC (area under the receiver operating characteristic curve, hereinafter AUC) of 0.92 and predict Not Helpful Status with an AUC of 0.97 (see the validation of model parameters in section S6.3).

Multidimensional ideology scaling of accounts on X

To examine the ideological scale inferred by X’s system, we rely on the method developed by Ramaciotti Morales et al. (35), which adapts Barberá’s (34) methodology to multidimensional settings. Specifically, we identify X accounts following Members of Parliament (MPs) from 23 countries across five continents, retaining those that followed at least three MPs and were followed by at least 25 accounts, following Barberá’s protocol (33–35). We then infer, through correspondence analysis, a multidimensional latent space arranging MPs and their followers by homophily (see details in section S5.1).

For the analysis of Community Notes latent ideology parameter, we consider the position of X accounts along δ₁, the first principal component of the correspondence analysis, which best explains which MPs they follow. Additionally, we map the positions of accounts onto interpretable political dimensions of the Global Party Survey (39), an expert survey positioning parties along several dimensions endowed with a reference frame making them comparable across countries, namely, (i) the Left-Right dimension, a leading dimension of polarization relevant for studies of misinformation in social media (40–42), and (ii) the Anti-Elite dimension, which measures criticism toward elites and institutions. In contrast to results regarding online misinformation in the United States (42), the degree of negative attitudes toward institutions has been shown to play a role in the probability of disseminating misinformation (36, 37). When comparing latent ideologies θ_n and θ_r with Left-Right and Anti-Elite dimensions, we focus on 13 countries where at least a thousand X accounts with inferred political positions have authored posts with proposed Community Notes. These countries represent a geographically diverse sample (see section S5.3): United States, United Kingdom, Japan, Spain, France, Brazil, Canada, Germany, Argentina, Israel, Australia, Poland, and Mexico.

RESULTS

Notes and ratings

Community Notes usage reveals highly unequal distributions (see section S2.1). Among the 1.1 million users self-enrolled in the program as of March 2025, 28.9% have rated fewer than 10 notes. Just 1% of the 405 thousand contributors with note-writing privileges authored 31.2% of all proposed Community Notes. Similarly, 35.8% of all proposed Community Notes are associated with posts published by only one thousand X accounts (see most commented accounts per country in section S3.5.1). Of the posts for which users requested Community Notes, 79.3% were flagged by only a single user, and a note was proposed for only 12.5% of these posts.

Figure 1A reveals the worldwide usage of Community Notes. The majority of notes relate to the United States, followed by Japan, the United Kingdom, Brazil, and France (see the country segmentation methodology in section S3.5). In the European Union, the number of Community Notes contributors per country correlates strongly with the number of active logged-in users (Pearson $\mathrm{\rho }=0.823,P<{10}^{-4}$), as disclosed under European transparency obligations.

Fig. 1. Community Notes usage and outcomes.

(A) Fraction of Community Notes proposed per country. To be associated with a country, a note must be predominantly evaluated by contributors who typically rate notes in that country’s language and that reference national websites (see section S3.5.1). (B) Cumulative distribution of time needed for a proposed note to reach Helpful Status via X’s algorithm, measured from the publication of the original post and from the publication of the note (truncated tails after 36 hours, representing 9.2 and 17.9% of notes, respectively). (C) Fraction of notes reaching Helpful Status in the United States (left) and in the United States when referencing expert fact-checks (right), segmented by the ideological leaning of note authors. Error bars show 95% CI via bootstrapping over notes and fact-checking organizations (see section S71.1).

Of Community Notes, 96.2% contain URLs pointing to supporting sources that aim to contextualize the original post, after preprocessing (19, 22). To examine news media citation practices in Community Notes, we compiled a list of 2134 global outlets from BBC Media Guides, supplemented by manual review of the most frequently referenced domains in notes. These established media outlets constitute the predominant sources in notes, appearing in at least 25.9% of notes, followed by X posts (17.4%, where contributors cite other posts as evidence) and Wikipedia articles (8.9% of notes).

Additionally, we curated a list of web domains from professional fact-checking organizations that are International Fact-Checking Network signatories or appear in Duke Reporter’s Lab directory, revealing that expert fact-checking articles appeared in only 3.5% of proposed notes (see the detailed curation of different types of sources in section S3.2).

Political content accounts for 65.2% of posts and 76.6% of Community Notes ratings, with sports as the second most frequent category (6.3% of posts) (see complete distribution and validation in section S3.3). To further characterize notes discussing politics, we use a topic model (43) trained on political party manifestos from the Manifesto Project (44). Among the topics included in this scheme, the ones most discussed in notes are Political Authority (25.6% of posts, including praise and criticism of specific parties), Law and Order (12.4%, including mention of crimes, the police, and the judiciary), Equality (7.2%, mentioning underprivileged groups, social classes, economic redistribution, and discrimination), Environmental Protection (6.6%, including natural resources, natural preservation, and animal rights), and Freedom and Human Rights (6.2%, mentioning freedom of speech, press and assembly) (see a detailed distribution and validation of political topics in section SC.4).

Last, we observe that, across analyzed countries, Community Notes are requested and proposed under posts authored by accounts throughout the political spectrum. Of the nondeleted posts that received proposed Community Notes and have authors with inferred ideological leanings, 58% are from right-leaning accounts in the United States, 54% in Japan, and 53% in France (see Left-Right leaning distribution per country in section S5.3).

Algorithmic resolution outcomes

We now examine the status assigned by X’s Community Notes algorithm to proposed notes. Of all proposed Community Notes arguing the original post is misinformed or potentially misleading, only 11.97% garner sufficient favorable ratings from “diverse” contributors to attain Helpful Status and be made visible on X. The remaining notes are not publicly displayed, either because consensus determined them to be Not Helpful (2.81%; hereinafter, achieving Not Helpful Status) or because they lacked sufficient ratings from diverse contributors to reach a conclusive status (85.22%). Of posts for which at least five users requested Community Notes, 17.1% ultimately received a proposed note that reached Helpful Status. Between January and March 2025, half of the notes that achieved Helpful Status did so within 4 hours 47 min from note submission and 15 hours 17 min after the original post publication; see distribution in Fig. 1B.

At the account level, we observe remarkable variability in the proportion of posts with notes achieving Helpful Status. Politically divisive accounts show fewer notes achieving Helpful Status, for example, those of Elon Musk or Kamala Harris’s campaign have 3% or less than 0.1% of their annotated posts receiving such notes, respectively. Conversely, parody, entertainment, and conspiracy theory accounts exhibit a high proportion of notes achieving Helpful Status (see lists of salient accounts and metrics in section S4.3). More generally, 9.7% of annotated posts authored by accounts that held blue check verification before X’s 2022 policy change and were deemed of “public interest” (45) received notes with Helpful Status, compared to 14.0% for unverified accounts (computations restricted to nondeleted posts).

Notes related to politically divisive issues struggle to garner consensus and achieve Helpful Status across each analyzed country. In particular, we consider four elections taking place during the period of observation of our study. Figure 2 shows a comparison between the fraction of notes reaching Helpful Status for election-related notes and for overall notes in the same period, for the 2024 US presidential (6.9 versus 14.7%), 2024 UK general (8.2 versus 10.9%), 2024 France legislative (12.8 versus 16.8%), and 2025 Germany federal elections (8.6 versus 11.9%). For each selected election, consensus is reached less frequently on election-related Community Notes compared to overall notes published during the same period. For comparison, 20.5% of notes reporting a scam and 17.1% of notes arguing a post contain manipulated media reached Helpful Status. See section S4.2 for a curation of keywords used to identify notes relating to elections and other topics.

Fig. 2. Comparison of the fraction of moderated notes.

Fraction of notes reaching Helpful Status per country during elections. The analysis compares election-related notes with the overall corpus (in the same period, ranging from 1 month before to 1 week after the election) during the 2024 US presidential, 2024 UK general, 2024 France legislative, and 2025 Germany federal elections. Error bars represent 95% confidence intervals, estimated by bootstrapping over the keywords used to identify election-related notes. Fraction of notes reaching Helpful Status per argued misleading type worldwide.

Last, as reported in Fig. 3C, posts with notes that reached Helpful Status are authored by accounts across the political spectrum in the 13 countries for which we computed ideological positions (see section SE.3 for Left-Right distributions across all countries). In the United States, 8.3% of annotated posts authored by Left-leaning users receive notes that achieve Helpful Status, compared to 12.9% for Right-leaning users, as observed in (46). However, this disparity should not be interpreted as definitive evidence that content from Right-leaning users is flagged more frequently. Because our analysis is restricted to nondeleted posts, the observed difference could be explained, at least partially, by differential deletion patterns, whereby Left-leaning users may delete their posts with Helpful Status notes more frequently than Right-leaning users (see discussion in section S7.1.1).

Fig. 3. Posts with Community Notes requested, written and achieving Helpful Status across ideology and counties.

Distribution of Left- and Right-leaning X accounts that authored posts for which (A) Community Notes were requested by users; (B) Community Notes were proposed by contributors; and (C) Community Notes reached Helpful Status. Statistics computed over nondeleted posts authored by accounts with an inferred ideological leaning (see section S5.1).

Latent ideological dimensions in Community Notes

In the United States, the single-dimensional latent ideological space learned by X’s Community Notes system closely aligns with users’ Left-Right political leanings: Notes with ${\mathrm{\theta }}_n>0$ appear under posts published by Left-leaning accounts, whereas notes with ${\mathrm{\theta }}_n<0$ are associated with Right-leaning accounts (Fig. 4A). Of notes associated with posts authored by Republican members of Congress, 95.0% have ${\mathrm{\theta }}_n<0$, whereas 94.6% of those associated with Democratic members of Congress have ${\mathrm{\theta }}_n>0$ (see the average latent ideologies θ_n of notes associated with key political figures across multiple countries in section S7.1). Contributors with ${\mathrm{\theta }}_r<0$ tend to rate notes as Helpful when they appear under Left-leaning accounts but rate notes as Not Helpful when they appear under Right-leaning accounts (Fig. 4B), and the opposite pattern holds for raters with ${\mathrm{\theta }}_r>0$ (Fig. 4C), aligning with the observations made by Allen et al. (17) during the pilot program. In the United States, the Left-Right political leaning of users predicts the sign of the latent ideologies θ_n of the notes associated with their posts, with an AUC of 0.822. Additionally, news outlets cited in Community Notes align ideologically with contributor leanings; we found a Pearson correlation of $\mathrm{\rho }=0.744$ ($P<{10}^{-4}$) between Ad Fontes Media bias ratings (47) and the latent ideology θ_r of contributors citing these media outlets as sources in their notes (see section S7.1.1).

Fig. 4. Ideological positions of notes according to the Community Notes algorithms on multidimensional national political spaces.

(A to C) Distribution of Left-Right leaning of X accounts (A) segmented by the sign of the community note latent ideology θ_n, (B) rated helpful/not-helpful by raters with negative latent ideology ${\mathrm{\theta }}_r<0$, (C) rated helpful/not-helpful by raters with positive latent ideology ${\mathrm{\theta }}_r>0$. (D to G) X accounts positioned in the (Left-Right; Anti-Elite) two-dimensional (2D) plane, segmented by the majority sign of the latent ideology θ_n of their associated Community Notes (red for ${\mathrm{\theta }}_n<0$ and blue for ${\mathrm{\theta }}_n>0$). The direction that best separates notes by majority outcome, learned through logistic regression, is displayed as CN along with the corresponding AUC score for sign prediction based on account positions in the 2D plane (SD from 10-fold cross-validation). The direction structuring each country’s political landscape δ₁ is projected onto the plane.

In the United States, expert fact-checking articles are used as sources by contributors across the ideological spectrum, although used predominantly ($60.3\%[58.0\%,62.8\%]$) by Left-leaning contributors (${\mathrm{\theta }}_r<0$). As displayed in Fig. 1C, Community Notes incorporating expert fact-checking sources attain Helpful Status at significantly higher rates ($21.3\%[18.3\%,24.1\%]$) than average, with this effect more pronounced for notes authored by contributors deemed Right-leaning according to the Community Notes model (${\mathrm{\theta }}_r>0$), $26.9\%[23.3\%,30.8\%]$ versus $17.6\%[14.0\%,20.7\%]$ for Left-leaning contributors [95% confidence intervals (CIs) obtained by bootstrapping over notes and fact-checking organization; see section S7.1.1].

The latent ideology computed by the Community Notes model (θ_n and θ_r) shows varying degrees of alignment with the Left-Right and the Anti-Elite dimensions computed with ideology scaling and political surveys. Figure 4 (D to G) maps X users with proposed Community Notes along the Left-Right and Anti-Elite dimensions in the United Kingdom, Japan, France, and Brazil, representing the geographical diversity covered by our study, including the Americas, Europe, and Asia. The reader is referred to fig. S31 for the rest of the countries, showing how CN captures the ideological structure across national settings. The latent dimension learned by X’s system (CN in Fig. 4), along which notes and raters are positioned, aligns closely with each country’s most structuring direction (δ₁ in Fig. 4) in this Left-Right/Anti-Elite plane. We quantify this alignment by measuring how well the position of a user along δ₁ predicts the sign of the latent ideology θ_n of notes associated with their posts. AUC values range from 0.850 in Poland to 0.729 in Israel, averaging $0.808\pm 0.037$ across 13 countries. Complementarily, user positions in the Left-Right/Anti-Elite plane predict the sign of the latent ideology θ_n of notes with an AUC of $0.813\pm 0.035$ across all 13 countries (see breakdown for all countries in our study in section S7.2).

We find that the direction learned by X’s system to structure content veracity disputes (direction CN in Fig. 4) aligns with the Left-Right dimension but cannot always be reduced to it, as anticipated by recent literature (36, 37). Instead, it may present additional alignments with attitudes toward elites and institutions (and, to a lesser degree, possibly additional issue dimensions not considered in our study), as highlighted by the cases of Japan and Brazil (Fig. 4, F and G).

DISCUSSION

We combined X’s Community Notes data, the algorithmic representation learned by this service, and ideological scaling for characterizing its operation in polarized settings. We showed, in 13 countries, that notes are requested, proposed, and displayed under posts authored by accounts throughout the political spectrum. The majority of proposed Community Notes discuss politics and predominantly rely on news outlet articles as sources, followed by X posts and Wikipedia articles.

The Community Notes system was originally designed for deployment among US audiences, with an architecture predicated on a single ideological dimension to be automatically inferred and that is considered as structuring of political polarization in the United States. Prior research has not systematically assessed the degree to which this foundational design principle determines the system’s effective functioning; specifically, the extent to which the ideological positions inferred for notes and raters correspond to ideological dimensions as measured by other means for large populations of users. Crucially, no prior work has examined how these assumptions and design principles translate to the operation of Community Notes systems worldwide, across countries characterized by distinct polarization structures. We address both questions through multidimensional ideology scaling across 13 countries. This study provides an empirical assessment of the extent to which the Community Notes system captures ideological alignment among US users and how this capacity generalizes across diverse national political contexts in Europe, Central and South America, and Asia.

X’s Community Notes system positions notes and raters along a learned latent ideological axis that we show closely aligns with each country’s most structuring political dimension. This latent axis corresponds to the Left-Right dimension and, with differences across countries, to an Anti-Elite dimension. These findings provide the first empirical support for the key hypothesis underlying X’s Community Notes’ design in a global comparative context.

Our results highlight shortcomings of this crowd-sourced system. While notes displayed on X (i.e., having reached Helpful Status) have been shown to align with professional fact-checking assessments (4, 19), factual but polarizing content may still fail to be resolved as Helpful. Only ¹/₁₀ of proposed notes ultimately appear on X, with even lower rates for notes associated with politically prominent figures, popular accounts, and divisive issues. More generally, and by design, the most polarizing content relating to the structuring dimension of national politics typically cannot be algorithmically resolved by Community Notes, regardless of their factuality. This limitation is especially relevant to events that activate such structural lines of division, such as national elections, pointing toward the potential complementarity between expert and crowd-sourced fact-checking, as originally intended by X (48). Furthermore, the fraction of polarizing content that will remain unresolved (regardless of factuality) is a design choice made by the platform, both explicitly by fixing the ideological diversity threshold for a note to reach Helpful Status and implicitly by adjusting the regularization strength of the various latent parameters of the Community Notes model.

We acknowledge limitations in our study. In particular, our analyses rely on a snapshot of the data produced by the service until 1 March 2025 and disregard potential fine-grained temporal dynamics. Additionally, our segmentation of notes and raters per country is based on their involvement in the national political discussions by selecting those that follow MPs, combined with their use of the country’s national language. Given the absence of additional metadata disclosed by the platform, we argue that this heuristic is reasonable and necessary for analyzing ideological dynamics in Community Notes.

Our post analysis was limited to content that remained available at the time of collection in March 2025, introducing a potential bias in the analysis as posts having notes achieving Helpful Status are more likely to be removed by their authors (26). Last, X’s Community Notes contributors are self-selected, and the platform provides no information to characterize this population demographically, let alone politically.

As platforms transition from expert fact-checking to crowd-sourced moderation, with YouTube (49), TikTok (50), and Meta (51) deploying similar systems, the latter explicitly using X’s open source algorithm as the basis of their rating system, this study provides a timely and comparative characterization of their operation in polarized settings across diverse countries. Several arguments exist for considering alternatives to expert fact-checking as these approaches accumulate challenges with regard to the required scale, reactivity, and efficiency (11) but also concerning perceived legitimacy and credibility (1). Yet, our results show that X’s Community Notes system, by relying on consensus rather than factuality, presents inherent limitations in addressing polarizing content. The use of crowd-sourced moderation as a substitute for expert fact-checking, therefore, warrants careful consideration and calls for risk assessment comparing the moderation outcomes of these approaches over contested polarizing content widely disseminated during events such as elections. While a direct comparison of claim-processing rates between expert-based and crowd-sourced fact-checking systems is methodologically difficult, as is the evaluation of the rates at which false or misleading content is identified, our results indicate that substituting the former with the latter introduces a demonstrable and structural asymmetry in the moderation of politically salient content, such as that related to elections, with respect to other types of claims, thereby generating a consequential risk.

MATERIALS AND METHODS

Ethics and compliance declarations

Our study did not involve experimentation with human subjects. All data used in our study are publicly available through X’s API and through the Community Notes program at https://x.com/i/communitynotes/download-data. The processing of political attributes data was declared on 19 March 2020 and 15 July 2021 at the registry of data processing at the Fondation Nationale de Sciences Politiques (Sciences Po) in accordance with General Data Protection Regulation 2016/679 and X policy. For further details and the respective legal notice, please visit the web page of project European Polarisation Observatory: medialab.sciencespo.fr/en/activities/epo/. The processing of the data has been approved by the Research Ethics Committee of the Paris Institute of Political Studies in its decision no. 2023-038.

Data collection

X publicly and continuously releases all Community Notes contributions enabling third-party scrutiny. As of 1 March 2025, 1.49 million X users have submitted 4.83 million requests for Community Notes across 1.91 million posts, 1.88 million notes were submitted and rated over 135.17 million times.

To comply with platform policies, user privacy rights, and deletion requests, X’s Community Notes dataset only discloses post IDs associated with notes, without further information on post content or author. To enable contextual analysis beyond ratings, we collected post content and author information, associated with those notes. As of March 2025, 616,104 X posts associated with 83.5% of nondeleted Community Notes were collected. The remaining posts were unavailable at the time of collection, i.e., having been removed. Additionally, of the 966,000 posts with at least two user requests for a Community Note, we collected 804,000 posts (83.2%); the remaining posts were unavailable at the time of collection in May 2025. It should be noted that such a collection process (as opposed to live stream collection or access to data reserved to the platform) introduces an inherent observational bias. Some analyses are limited to posts that remain available at collection time, which then captures the tendency of users to remove their posts after the note reached Helpful Status. This observational bias should be considered when interpreting the following results. Analyses that rely solely on the content of Community Notes, however, such as source analysis, are not marred by such limitation.

Topic analysis

To classify community note topics, we used a two-tiered approach. We first translated all posts to English using Facebook’s M2M100 translation model (52), ensuring consistent classification performance across original languages. For general topic classification, we used topic taxonomy and model by Antypas et al. (53), selected for its open-weight architectures and top-tier performance. For fine-grained political classification, we used Manifestoberta (43), a language model based on XLM-RoBERTa (54), trained on 1.6 million annotated statements from the Manifesto Corpus (55) following the Manifesto Project coding scheme (44). We report validation results for both approaches in the Supplementary Materials. Last, using a curated set of keywords listed in the Supplementary Materials, we extracted community notes related to elections held in the United States (2024 presidential election), United Kingdom (2024 general election), France (2024 legislative election), and Germany (2025 federal elections).

Latent ideology

We learn biases ${\mathrm{\beta }}_0,{\mathrm{\beta }}_r, \text{and} {\mathrm{\beta }}_n$ and latent ideologies ${\mathrm{\theta }}_r \text{and} {\mathrm{\theta }}_n$ for notes and raters from the publicly released community notes and ratings. We replicate X’s approach (19, 22) by implementing the matrix factorization of the rating model ; computational details and validation are provided in the Supplementary Materials (Section S6).

To interpret the learned latent ideology dimension in terms of Left-Right and Anti-Elite positions, we use the ideological scaling of X users developed by Ramaciotti et al. (35). This approach extends Barberá’s (34) methodology to identify multidimensional ideological positions of users using political survey data across countries. We collected X users following MPs across 23 countries: France, Germany, United States, United Kingdom, Spain, Brazil, Australia, Italy, Israel, Argentina, Denmark, Sweden, Canada, Poland, the Netherlands, Switzerland, South Africa, Nigeria, Finland, Mexico, New Zealand, Japan, and Belgium. We retained only users following at least three MPs and followed by at least 25 accounts, following Barberá’s protocol (29, 33, 34). We then estimate ideal positions for users and MPs using correspondence analysis (56) on the follower-MP adjacency matrix. Validation and interpretation are reported in the Supplementary Materials.

Supplementary Materials

This PDF file includes:

Sections S1 to S7

Figs. S1 to S32

Tables S1 and S2

References