Community Notes on

---

Team members

Facilitators: Olesya Razuvayevskaya and Richard Rogers

Design facilitator: Alessandra Facchin

Participants: Ulrikke Dybdal Sørensen, Bumju Jung, Maricarmen Rodríguez Guillen, Wilma Ewerhart, and Adel Tayebi

---

Links

For a quick overview, please refer to the poster with the main outcomes of the project:

A State of Non-Agreement: Community Notes on X

Summary of Key Finding

Our analysis indicates that Community Notes (CNs) tend to foster dissensus rather than consensus, both at any given moment and in their final status. In other words, users frequently disagree on whether a note is helpful or unhelpful, regardless of whether it addresses misinformation or non-misinformation. Regarding timeliness, CNs are typically published well after the peak of post dissemination which often happens within the first five hours of spread (Renault et al, 2024), with an average delay of four days. Contrary to the common assumption that crowdsourced moderation is faster than professional fact-checking, the mean time lag between a post on X and the corresponding CN publication is 19 hours, which falls within the typical range for professional fact-checkers (10–20 hours, Shao and Ciampaglia, 2016). Furthermore, CN authorship is highly skewed: 80% of all notes are contributed by fewer than 27% of authors. Notably, high output does not equate to high quality—authors producing the most notes tend to generate significantly fewer helpful notes, on average, than those who contribute less frequently.

1. Introduction

Online content moderation for misleading information typically relies on either automated detection or one of two manual approaches: (i) expert fact-checking or (ii) crowd-sourced moderation. As the volume of false information continues to grow, manual expert fact-checking has become increasingly intractable, requiring expertise across multiple languages and region-specific domains (Martel et al., 2024; Stencel et al., 2024). Moreover, it carries risks of bias, censorship, and partisanship stemming from the judgments of a small group of experts (Flamini, 2019). In an effort to address scalability and bias concerns, platforms have increasingly turned to crowdsourcing as an alternative to expert fact-checking. This shift is exemplified by Meta’s announcement in early 2025 that it would discontinue its fact-checking program in favor of Community Notes, a system already deployed by X (formerly Twitter) under its earlier name, Birdwatch.

The move toward crowdsourced moderation has been interpreted in different ways: as aligning with prevailing political trends—described by one technology industry columnist as a “surrender to the right on speech” (Newton, 2025)—and as adopting a new moderation philosophy grounded in consensus-building rather than adjudication or expert authority (de Keulenaar, 2025). However, CNs are still very reliant on fact-checking sources, with around 33% of notes reported to cite at least one such source (Borenstein et al., 2025). Moreover, they can be quite vulnerable to coordinated attacks, potentially even suppressing helpful notes by downvoting them (Augenstein et al., 2025). This raises a central question: how effective is the Community Notes approach?

In this study, we attempt to answer this question by investigating three primary research questions. First, we assess the extent to which CNs foster consensus or dissensus. Second, we examine the timeliness of CN publication, given that “noting” a post can influence its reach either by reducing its visibility or prompting its removal. Third, we analyze the composition of the crowd to determine whether a small subset of users generates the majority of notes, shedding light on potential concentration of influence and bias—issues that crowdsourced fact-checking is often assumed to mitigate. Finally, we extend our analysis to examine bias more directly by identifying the propaganda techniques used by note authors and evaluating the overall quality of the notes they produce.

2. Dataset

This study draws upon the publicly available, open-source dataset of Community Notes (CNs) daily updates provided by X. The dataset comprises five principal categories, of which four were utilized in our analysis: (i) Community Notes data, containing the full text of each note, the associated post identifier, and the categorical labels assigned when flagging a post; (ii) Note rating data, which, when aggregated, enables the determination of a note’s perceived helpfulness; (iii) Note status history, documenting the sequence of status changes from initial creation—when the note is first designated as “Needs more ratings”—to subsequent classifications as “helpful” or “unhelpful,” and ultimately to a permanently locked state two weeks after creation; and (iv) User enrollment status data, which provides information on the registration and participation status of contributors to the CN program. We retrieved all available records on 1 July 2025, resulting in a corpus of approximately 2.1 million notes.

3. Research Questions

RQ1: How effective are Community Notes in building consensus or dissensus?

To assess the Community Notes (CN) system’s capacity to foster agreement, we examine whether it achieves its stated design goal—building consensus—or whether it more often results in sustained disagreement (dissensus). This is addressed through the following sub-questions:

• RQ1.a: Do Community Notes tend to build consensus?

• RQ1.b: Why are Notes in a state of (non-)agreement?

Although CNs are explicitly designed to facilitate agreement among contributors, prior scholarship and industry reports on user-driven content moderation have raised doubts about whether such systems are as rapid, scalable, and transparent as claimed (e.g. Braga et al. 2025). In this context, we investigate whether consensus is in fact the dominant outcome (RQ1.a) or whether the system’s dynamics tend instead toward dissensus, and if so, why (RQ1.b).

RQ2: How timely and qualitatively effective are Community Notes?

The effectiveness of CNs cannot be evaluated solely on whether they reach agreement; when this agreement is reached is equally critical. Timeliness is particularly relevant in the context of misinformation, where interventions are most impactful during the early stages of content diffusion. If CNs are posted after a post’s peak virality, their capacity to reduce its reach, prompt deletions, or shape public perception diminishes sharply.

To address this, we examine:

• RQ2.a: What is the average and median time lag between the original post and the publication of the corresponding CN?

• RQ2.b: Are notes rated as highly helpful associated with identifiable properties such as the quality of cited sources, the use of unbiased language, or the relevance of the note to the original claim, in comparison to less helpful notes?

This question integrates both temporal and qualitative dimensions of effectiveness, allowing us to determine whether CNs that arrive earlier—and with higher assessed quality—are more likely to fulfil their intended corrective function.

RQ3: Who constitutes the “crowd” in Community Notes, and how is participation distributed?

Crowdsourced moderation is often justified on the grounds that a broad and diverse participant base can reduce the biases of centralised expert fact-checking. However, if CN authorship is dominated by a small minority, the supposed benefits of diversity may be undermined, potentially reintroducing concentrated biases and limiting representativeness. Furthermore, over-reliance on a small cohort could negatively affect quality if prolific contributors consistently produce less helpful notes.

To investigate the structure and diversity of participation, we ask:

• RQ3.a: What percentage of authors are responsible for 80% of CNs?

• RQ3.b: Does the majority of authors produce helpful notes?

• RQ3.c: Do the most frequent contributors tend to produce helpful notes?

• RQ3.d: Which user enrolment statuses are associated with the majority of published notes?

• RQ3.e: What are the topic and persuasion techniques dynamic across the CN platform?

By addressing these questions, we can evaluate whether CNs are genuinely “crowdsourced” in practice, or whether they function more as a specialised moderation activity dominated by a narrow set of contributors.

4. Methodology

4.1 How effective are Community Notes in building consensus or dissensus?

Similar to Wikipedia and Reddit’s models of community-driven moderation, the CN system on X relies on the voluntary contributions of its users and the attainment of consensus to evaluate content posted on the platform (Wirtschafter & Majumder, 2023). In CN, consensus is operationalised as the achievement of what the platform terms a “Diversity of Perspectives.” This metric is computed using a bridging algorithm that determines whether a given note’s ratings come from individuals whose prior rating behaviours indicate differing perspectives (X; Crisan & McNutt, 2025). Crucially, X does not incorporate any explicit demographic, political, or experiential attributes of contributors when inferring these perspectives; rather, “diversity” is derived entirely from statistical patterns in how contributors have rated other notes.

This process constitutes what can be described as a manufactured or computational consensus—a consensus defined not by direct deliberation or explicit representation, but by an inferred “viewpoint space” constructed through algorithmic classification. As Kovermann (2025) observes, such an arrangement means that consensus in CN is mediated by algorithmic design choices, contributor eligibility rules, and statistical agreement thresholds. In practice, a note is considered to have reached consensus when it receives a sufficient number of ratings classifying it as “helpful,” and these ratings are distributed across algorithmically inferred diverse perspectives.

We operationalise consensus in the Community Notes (CN) system as the state in which contributors collectively agree to either publish a note or reject it. Consensus may thus be reached for both published and unpublished notes. In the NoteStatusHistory dataset, this corresponds to the statuses CURRENTLY_RATED_HELPFUL or CURRENTLY_RATED_NOT_HELPFUL. A note is classified as consensus and published when its current status is CURRENTLY_RATED_HELPFUL and its classification in the Notes dataset is MISINFORMEDORPOTENTIALLY_MISLEADING. Conversely, a note is classified as consensus but unpublished when its current status is CURRENTLY_RATED_NOT_HELPFUL, with its classification being either MISINFORMEDORPOTENTIALLY_MISLEADING or NOT_MISLEADING.

We operationalise dissensus as the absence of such agreement. This is represented in the NoteStatusHistory dataset by the status NEEDS_MORE_RATINGS, which indicates that a note has not yet met the algorithmic consensus criterion of “Diversity of Perspectives” and therefore has not been agreed upon as either helpful or unhelpful. While the CN system is explicitly designed to foster consensus, the persistence of dissensus can emerge as an unintended outcome, potentially reflecting user misuse, trolling behaviour, or strategic non-cooperation. However, dissensus is not necessarily accidental: in some cases, it may be deliberately evoked, promoted, or sustained as part of ongoing contestation over a note’s validity. In this sense, dissensus in CN should also be understood as a state of non-agreement that is continuously subject to negotiation, shaped by the system’s design and the social dynamics of its contributors.

One way to investigate the underlying causes of dissensus in the Community Notes (CN) system is to examine the concept of consensus itself and the theoretical implications of its absence. Drawing on deliberative democratic theory, consensus is often linked to ideals of collective reflection, democratic agreement, and open dialogue between diverse perspectives—what Kovermann (2025) refers to as deliberative consensus or open deliberation. Within this framework, a state of non-agreement need not simply be interpreted as system failure; it can also be seen as a discursive space for contestation, reflection, and negotiation of meaning.

Previous studies have shown that CN contributors, in their practices of “noting” or rating, do not exclusively engage in content moderation by flagging misinformation or providing clarifying context. They also use the system to comment on and debate the notes themselves. This became more explicit in 2023, when X introduced the rating category NotHelpful _NoteNotNeeded (NNN). In the note creation interface, authors can also select this category and provide a short summary. These summaries often contain explanations, elaborations, or arguments—transforming the NNN label into a vehicle for commentary rather than simply a binary judgement of helpfulness.

In this way, NNN facilitates a form of non-algorithmic engagement within the CN system: rather than directly contributing to the computational consensus model (which relies on the “Diversity of Perspectives” metric), participants use the platform to exchange arguments, counterpoints, and meta-commentary. This mode of interaction aligns more closely with a deliberative model of consensus-building, where dissensus is not merely a breakdown of agreement but part of an ongoing democratic dialogue.

To examine the role of NNN in shaping dissensus, we conducted two complementary analyses. First, we identified the three posts that had received the highest number of notes overall. For each of these posts, we examined the top note in terms of the number of ratings it received, analysing the distribution of rating categories. This allowed us to determine whether highly engaged notes served primarily as vehicles for debate rather than as consensus-building tools in line with CN’s stated purpose—namely, to “add helpful context to posts and keep people better informed.”

Second, we extracted all notes containing “NNN” in their summary field and analysed them qualitatively to identify whether they were used to comment on or critique other notes, or to dispute the premise or framing of the note in question. This enabled us to assess whether NNN functions as a form of dissent and debate within the CN community, and whether such practices contribute to the persistence of dissensus in the system.

4.2 How timely and qualitatively effective are Community Notes?

We operationally define the effectiveness of a Community Note as its publication following a review process in which it is deemed sufficiently helpful by other note contributors. To approximate the publication time of a note, we utilize the Note Status History dataset, which records several relevant timestamps:

• The first instance when the note’s status transitioned from “Needs More Ratings” (NMR) to either “Helpful” or “Not Helpful.”
• The timestamp corresponding to the change to the current status (either “Helpful” or “Not Helpful”).
• The timestamp of the most recent non-NMR status update.
• The timestamp when the note was locked, i.e., when its status became permanent as either “Helpful”, NMR or “Not Helpful.”

A two-step conditional process designed to ensure that only notes that have reached a final, stable state are considered, and that the earliest possible time for that state is captured.

Step 1: The Master Condition (Gatekeeper Logic) The first and most critical step is to check a note's lockedStatus. This acts as a "gatekeeper" to determine if a note is eligible for time calculation at all. For a "successful" note: The script only considers notes where the lockedStatus is exactly CURRENTLY_RATED_HELPFUL. If the lockedStatus is anything else (e.g., NEEDS_MORE_RATINGS, NOT_HELPFUL, or empty), the note is immediately excluded from the "Helpful" time analysis. For a "Not Helpful" note: Similarly, the script only considers notes where the lockedStatus is exactly CURRENTLY_RATED_NOT_HELPFUL. This initial check is crucial because lockedStatus represents the note's final, settled state after all rating activity has concluded.

Step 2: The Earliest Timestamp Search Once a note passes the "Master Condition" (e.g., its lockedStatus is confirmed as CURRENTLY_RATED_HELPFUL), the script then performs a comprehensive search to find the absolute earliest moment that "Helpful" status was ever recorded:

Notes that never attain the “Helpful” status at any recorded stage are excluded from this analysis. It should be noted that the analysis is constrained by data availability and structure. Specifically, it is hypothetically possible for a note to have a first, current, and most recent non-NMR status all marked as “Unhelpful,” while having been rated “Helpful” at some intermediate point not captured in the dataset due to incomplete status tracking. This limitation introduces a potential source of error which we accept for the purposes of this study.

The timestamps are recorded in milliseconds since the epoch and were converted to Coordinated Universal Time (UTC). We linked the Note Status History dataset with the Notes dataset through the note ID to extract the corresponding Tweet ID. Using Python’s Snowflake library, we retrieved the exact posting time of each tweet. The time lag between the tweet posting and note publication was calculated by subtracting the note publication timestamp from the tweet’s posting timestamp. Both the mean and median time lags were computed to provide a robust estimate; median values are emphasized to mitigate the influence of outliers, such as delayed virality of older tweets.

On January 20, 2023, the platform implemented a retrospective lock on all community notes, after which note statuses became immutable. Additionally, a two-week timeline was introduced, whereby a note’s status becomes permanently locked two weeks after creation. Given that this policy significantly affects the dynamics of note status changes, our analysis restricts the dataset to notes published after this date, thus ensuring consistency and comparability of time-lag statistics. To improve data accuracy, the end of May 2025 was set as the data cutoff date, as the dataset only contained partial results for notes published in June 2025.

To investigate the relationship between note characteristics and their perceived helpfulness, we utilized the Note Ratings dataset. When evaluating a note, raters can specify the reasons they consider it “Helpful” or “Unhelpful.” Our analysis focuses on five specific “Helpful” attributes:

1. Clarity of expression.

2. Provision of important contextual information.

3. Use of unbiased language.

4. Inclusion of credible sources.

5. Direct engagement with the claim in the original post.

For each of these five attributes, we assessed whether its presence was associated with the highest level of helpfulness attained by a note. To this end, we conducted a chi-squared significance testing (Greenwood and Nikulin, 1996) , which evaluates whether the distribution of helpfulness levels is significantly different for notes possessing a given attribute compared to those without it. This approach allows us to quantify the extent to which specific note properties contribute to the likelihood of a note being rated as highly helpful.

4.3 Who constitutes the “crowd” in Community Notes, and how is participation distributed?

To estimate the concentration of the notes within a small group of note authors, we analyzed the Notes dataset by grouping notes according to the unique author identifier (noteAuthorParticipantId) and calculating the number of notes per author. We then applied Pareto analysis (Powell and Sammut-Bonnici, 2014) to assess whether note production follows an 80/20 distribution. The Pareto principle, originally observed by Vilfredo Pareto in the context of wealth distribution—where 80% of Italy’s wealth was held by 20% of its population—has since been found to characterize a range of systems with disproportionate input-output distributions.

To evaluate whether high-output authors tend to produce high-quality notes, we operationalized quality as the percentage of an author’s notes that were rated “Currently Rated Helpful” (currentStatus = CURRENTLY_RATED_HELPFUL). For each author, we computed:

This allowed us to determine whether prolific contributors also maintain a high proportion of helpful content. Using the Note Status History dataset, we further calculated the cumulative distribution of authors by the proportion of their notes in each of the three status categories: (i) Needs More Ratings (NMR), (ii) Helpful, and (iii) Not Helpful.

To specifically test whether the most prolific authors tend to produce helpful notes, we compared two groups:

• Top 10% contributors (authors with the highest note counts)

• Bottom 90% contributors (remaining authors)

For each group, we computed the percentage of authors meeting a 30% helpfulness threshold—defined as having at least 30% of their notes currently rated as Helpful.

We next examined whether author user status is associated with the number of notes authored. We linked the User Enrollment Status dataset with the Community Notes dataset using the unique author ID, grouping authors into five categories:

1. atRisk – Authors at risk of being locked if they receive one more unhelpful rating.

2. earnedIn – Active authors who have earned the right to post notes.

3. removed – Authors who have been removed or locked.

4. newUser – Registered users not yet permitted to post notes.

5. earnedOutAcknowledged – Locked authors who are attempting to regain posting privileges by producing helpful notes.

We then analyzed the distribution of authors across these categories and the total number of notes they contributed. This complementary analysis allowed us to estimate whether users with posting privileges frequently remain inactive, and whether the majority of active note authors are in statuses that place them at risk of locking.

Finally, we examined the thematic and rhetorical characteristics of notes authored by the top 1,000 contributors. To identify prevalent topics, we applied the multilingual classification model developed by the University of Sheffield (Wu et al., 2023) , which categorizes text into the following domains:

• Economy and Resources – Issues relating to financial costs and benefits, resource availability (physical, human, or financial), and the capacity of current systems to meet demands.

• Religious, Ethical, and Cultural – Considerations involving religious or ethical principles, as well as traditions, customs, and values connected to a given policy or social issue.

• Fairness, Equality, and Rights – Questions surrounding the equitable distribution of rights, responsibilities, and resources within society.

• Law and Justice System – Matters concerning the legal rights, freedoms, and powers of individuals, corporations, and government entities.

• Crime and Punishment – Discussions of the effectiveness, fairness, and broader implications of laws, law enforcement, and judicial processes.

• Security, Defense, and Well-being – Threats and opportunities affecting the safety, welfare, economic stability, and overall well-being of individuals, communities, or nations.

• Health and Safety – Topics related to healthcare, sanitation, disease prevention, and public safety measures.

• Politics – Political processes, actors, and agendas, including elections, lobbying, policy-making, public opinion, and demographic considerations.

• International Relations – Issues concerning a nation’s foreign policy, diplomatic relations, and international reputation.

For the same authors, we analyzed the distribution of persuasion techniques within each user status category (atRisk, earnedIn, and earnedOutAcknowledged). Our working hypothesis was that authors who avoid manipulative rhetorical strategies are less likely to be locked or at risk of being locked. To detect these strategies, we again used the multilingual classifier developed by the University of Sheffield (Wu et al., 2023) , which performs multiclass classification into the following propaganda types:

• Doubt – Undermining credibility by questioning an individual’s character or attributes rather than addressing the actual issue or evidence.

• Repetition – Reinforcing a point by reiterating the same idea multiple times within the discourse.

• Appeal to Fear–Prejudice – Persuading or rejecting an idea by exploiting fear, aversion, or pre-existing biases toward the idea or its alternatives.

• Appeal to Authority – Supporting a claim solely on the basis that it comes from a recognised authority, without presenting independent justification.

• Slogans – Using short, memorable phrases that often rely on stereotyping or emotional triggers to influence opinion.

• Loaded Language – Employing emotionally charged words or expressions to sway opinion rather than relying on factual reasoning.

• False Dilemma / No Choice – Presenting a situation as if there are only two possible options when, in reality, more alternatives exist.

• Flag-Waving – Promoting an idea by appealing to group pride or emphasising benefits to a specific community or identity.

• Name-Calling (Labeling) – Assigning derogatory or simplistic labels to individuals or groups to discredit them.

• Causal Oversimplification – Attributing an outcome to a single cause when multiple contributing factors exist.

• Appeal to Hypocrisy – Dismissing an argument by accusing its proponent of hypocrisy rather than engaging with the substance of the claim.

• Obfuscation / Vagueness / Confusion – Using ambiguous or unclear language that allows multiple interpretations, obscuring the main argument.

• Exaggeration / Minimisation – Overstating or downplaying the significance of an event, fact, or claim to distort perception.

• Red Herring – Distracting from the main topic by introducing unrelated or tangential issues.

• Guilt by Association – Discrediting a person or idea by linking it to a group or concept with negative connotations.

• Conversation Killer – Using statements that shut down discussion or discourage further critical engagement.

• Appeal to Popularity – Claiming an idea is valid because it is widely accepted or unchallenged.

• Straw Man – Misrepresenting an opposing argument in a weakened or distorted form in order to refute it more easily.

• Questioning the Reputation – Attacking the moral character or trustworthiness of a person to undermine their position.

• Appeal to Time – Arguing that an action is necessary because the “right time” has arrived, without deeper justification.

• Whataboutism – Responding to criticism by making a counter-accusation of hypocrisy, avoiding direct engagement with the original argument.

• Appeal to Values – Framing an idea as desirable by associating it with values the audience perceives as inherently positive.

5. Findings

RQ1.a: Do Community Notes tend to build consensus?

Our findings show that Notes are particularly dissensus building (Figure 5.1). Of all Community Notes written since January 2023, raters reached a consensus on 11.5% of them. 72% of these were published. The last 28% were left unpublished i.e. agreed to not being helpful. This shows that the CN system at the given time we retrieved the data, is particular consensus building when notes are helpful and detect misleading posts on X.

Figure 5.1 Treemap dendrogram illustrating consensus and dissensus in note ratings. Blue segments indicate consensus regarding a note’s classification as either Helpful or Not Helpful, whereas red segments represent the proportion of notes for which raters did not reach consensus. The dark blue segment specifically denotes the proportion of notes that achieved consensus, were rated Helpful by the community, and flagged the original post as misleading—thereby qualifying for publication.

For the other 88.5% of the Notes, there was dissensus as they ‘needed more ratings’, meaning that the state of the Community Notes system at the given time we retrieved data is dissensus building rather than consensus building.

The composition of consensus and dissensus corresponds as there is particular consensus and dissensus on the same misleading or not misleading sub-categories. Notes that are most agreed upon are about factual errors and missing important context while at the same time being the most disagreed upon.

RQ1.b: Why are Notes in a state of non-agreement?

The Community Notes (CN) system appears to foster conditions of dissensus rather than consensus, suggesting that the underlying concept of computational consensus on which the system is predicated may not be achieving its intended purpose. What, then, drives this form of dissensus? Our analysis indicates that the most heavily annotated posts on X tend to result in dissensus. Among the minority of notes that did reach consensus, none flagged the associated post as misinformation, implying that none of the notes attached to the top posts were ultimately published. Notably, two of the most highly rated notes—each exceeding 2,000 ratings, far above the five-rating minimum—remained in dissensus. The most frequent rating category for these notes was notHelpfulNoteNotNeeded, indicating that contributors and raters largely agreed that the note should not have been created in the first place. This dynamic points to a form of non-algorithmic engagement within the community, whereby participants interact primarily to express disagreement over a note’s necessity rather than to evaluate the accuracy or misleading nature of the underlying post.

Figure 5.2 The three posts that received the highest number of community notes as of 1 July 2025. Circles positioned above each post indicate the set of notes associated with that post. Circle size corresponds to the number of ratings each note received, while color denotes rating consensus: blue represents notes that achieved consensus, and red represents notes where dissensus persisted.

The “NNN” Phenomenon. Within Community Notes, the NoteNotNeeded label frequently appears in abbreviated form as “NNN.” While the intended function of “NNN” is to indicate that a note is unnecessary, our analysis reveals that contributors employ the label for a broader range of purposes. Examples include meta-commentary such as “Your suggested community note is best added as a response in the thread, not a note. NNN.” The most common rationale is that the note is perceived as inapplicable, often because the associated post constitutes a personal opinion, satire, a joke, or an unverified rumor (e.g., “NNN – don’t abuse community notes”). In other instances, contributors use “NNN” as a means of direct interpersonal interaction, either to comment on other notes or to address fellow contributors (e.g., “I’m replying to the person that wrote NNN”).

Noting as Debating. This broader use of “NNN” suggests that, rather than simply rating a note, some contributors engage in discursive exchanges within the rating system itself. Such practices transform Community Notes into a space for deliberation, rather than its stated function of facilitating consensus—whether genuine or manufactured—around information accuracy. This “noting as debating” dynamic can therefore be seen as a form of system repurposing, where participants use a tool intended for content evaluation as a forum for interaction and contestation. While this constitutes a “misuse” from the perspective of platform design and policy, it may also indicate an unmet user demand for structured discussion spaces within the platform ecosystem.

RQ2.a What is the average and median time lag between the original post and the publication of the corresponding CN?

Our analysis indicates that, on average, community notes are published 132.4 hours (Figure 5.4)—approximately more than five days—after the corresponding post. The median publication time is substantially shorter, with 50% of notes appearing within 9.5 hours. Previous research (Renault et al, 2024) indicates that the receipt of a Community Note substantially increases the likelihood that the associated tweet will be deleted. In our dataset, 23.4% of tweets identified by users as misinformation, and 9.2% of tweets not classified as misinformation, were deleted after receiving a note, excluding tweets from suspended accounts. However, prior work suggests that this intervention may occur too late to meaningfully mitigate the initial viral spread of misleading content. Tweet diffusion is rapid; approximately half of all retweets occur within the first five hours of posting (Renault et al, 2024).

 

On average, notes achieved a Currently Rated Helpful (CRH) status 21.7 hours after creation, with a median time of 6.2 hours. In contrast, the mean time to reach a Currently Rated Not Helpful (CRNH) status was 43.7 hours, with a more rapid median of 5 hours (Figure 5.3). The analysis of the time taken for a tweet to get the note and reach a "helpful" or "not helpful" status reveals distinct patterns. On average, it takes approximately 129.14 hours (about 5.38 days) for a tweet to achieve a "helpful" status, with a median time of 15.59 hours. For notes reaching a "not helpful" status of notes, the average time from tweet creation is slightly longer at approximately 135.13 hours (about 5.63 days), with a median time of 18.69 hours.  

Future investigations into the success rate of Community Notes (defined by "helpful status") should consider the impact of both tweet publication time and the delay in note creation. Preliminary findings suggest that tweets published in the afternoon are more likely to achieve helpful status. Additionally, a shorter delay between tweet publication and note creation appears to correlate with a higher frequency of helpful notes (Figure 5.5). Figure 5.6 suggests that while a greater number of notes may be created with delay, the success time to reach consensus is more dependent on less delay in note creation.

Figure 5.3. Consensus Attainment Time by Month. January 2023 - May 2025

Figure 5.4. Time delay between the publication of the tweet and note by Month. January 2023 - May 2025

Figure 5.5. Impact of time of creation of tweets and delay in creation of notes in the success time and frequency. January 2023 - May 2025

Figure 5.6. Effect of delay in creation of notes to the final status. January 2023 - May 2025

On 21 January 2023, Twitter/X revised its Community Notes policy, reducing the rating window for a note to two weeks, after which its status became fixed. Following this change, the proportion of notes achieving a “Currently Rated Helpful” (CRH) designation increased from 6.6% to 11.5%. The average time for a note to attain CRH status decreased substantially from 8,307.7 hours to 21.7 hours (Figure 5.3), while the median time fell to 6.2 hours, indicating that half of all CRH-designated notes reached this status in approximately six hours.

Consistent with prior research (Basset, 2025; Nenno, 2025) our findings indicate that Community Notes authoring and rating activity tends to peak around major political events. Following the October 7th incident in Gaza, the number of notes rose sharply to 61,881 in October 2023—nearly double the volume recorded in the preceding month. The most pronounced surge occurred in the lead-up to the U.S. presidential election, reaching 107,629 notes in July 2024, coinciding with the assassination attempt on Donald Trump. Elevated activity persisted through December 2024 (over 100,000 notes) and peaked again in January 2025 at 118,714 notes, during the post-election period. In recent months, however, note creation has shown a marked decline, suggesting a reversion toward baseline levels following the resolution of major political events.

RQ2.b Are notes rated as highly helpful associated with identifiable properties such as the quality of cited sources, the use of unbiased language, or the relevance of the note to the original claim, in comparison to less helpful notes?

Among notes that were accepted as helpful, 59.8% achieved the highest possible helpfulness rating. Statistical analysis reveals that notes are significantly more likely to be rated as highly helpful when they exhibit certain key qualities: clarity, the inclusion of credible sources, direct engagement with the claim in question, the provision of important contextual information, and the use of unbiased language (p < 0.005 for each attribute). These findings underscore the multifaceted nature of what constitutes an effective corrective note within crowdsourced fact-checking environments.

The significance of clarity suggests that comprehensibility is essential for community members to recognize and endorse the note’s value. Similarly, the presence of credible sources likely enhances the note’s perceived authority and trustworthiness, reinforcing its corrective potential. Addressing the claim directly ensures relevance and focus, preventing dilution of the note’s corrective purpose. The addition of important context supports a deeper understanding of the original post’s content, which may aid in correcting misinformation more effectively. Finally, unbiased language contributes to the note’s perceived neutrality, potentially reducing partisan resistance and fostering broader acceptance.

Collectively, these results align with broader literature on misinformation correction, which highlights the importance of transparent, evidence-based, and respectfully framed interventions to facilitate consensus and reduce resistance in online communities (Lewandowsky et al., 2012; Vraga & Bode, 2017). Consequently, these findings have practical implications for the design and moderation of crowdsourced fact-checking platforms, suggesting that guidelines and training for note authors could emphasize these qualities to enhance the overall effectiveness of the system.

RQ3.a What percentage of authors have authored 80% of the notes?

Our analysis reveals that 27.0% of note authors are responsible for producing 80% of all notes (see Figure 5.7). This finding exemplifies a common pattern observed in many online collaborative platforms, where a relatively small subset of highly active contributors generates the majority of content—a phenomenon often described by the Pareto distribution or the “80/20 rule.” In the context of Community Notes, this distribution highlights a pronounced concentration of participation, indicating a significant imbalance in user contributions. Such skewed engagement dynamics have important implications for platform governance and content diversity, as they suggest that the platform’s discourse and moderation efforts may be disproportionately shaped by a limited group of users.

Contrary to prior assumptions that community-driven platforms inherently promote broad participation, this concentration of activity aligns Community Notes with traditional fact-checking models discussed in Section 1, which have been critiqued for risks of censorship and partisan bias arising from the judgments of a limited group of experts (Flamini, 2019). This parallel raises important questions about the extent to which crowdsourced moderation systems like Community Notes may reproduce similar dynamics of centralized influence despite their participatory design. Understanding this imbalance is crucial for designing interventions aimed at fostering broader and more equitable participation in collaborative content creation.

Figure 5.7. Distribution of note authorship. The graph shows a cumulative distribution of authors vs. notes written. 1 July, 2025.

RQ3.b Does the majority of authors produce helpful notes?

Our analysis (Figure 5.8) reveals that 77.3% of note authors have 0.0% of their notes rated as Helpful, indicating that a substantial majority have not produced any content deemed helpful by the platform’s community of raters. The average helpfulness rate among all note authors is 6.8%, as marked by the vertical dashed line. This distribution demonstrates that most contributors receive low or no helpfulness recognition, with meaningful contributions concentrated within a relatively small subset of users. Such a pattern underscores a pronounced participation inequality in terms of quality contributions.

Similarly, the results show that 77.4% of authors have 0.0% of their notes classified as Not Helpful, suggesting that most contributors do not receive explicit negative evaluations. The average rate of Not Helpful notes across all users is 7.2%, as indicated by the mean marker. This finding likely reflects low engagement from raters rather than uniformly high-quality note production. Only a minority of authors generate a significant proportion of notes rated as Not Helpful, suggesting that problematic content remains limited within the community.

Perhaps most striking is the prevalence of notes categorized as “Needs More Ratings.” Here, 60.9% of contributors have 100% of their notes in this undecided status, with the overall average standing at 86.0%. This underscores a critical issue of rater engagement, where the majority of notes remain unresolved due to insufficient evaluations. This dynamic highlights that limited interaction, rather than solely content quality, substantially influences whether an author’s contributions achieve visibility and impact within the platform.

Figure 5.8. Distribution of note outcomes by status. The graph shows a cumulative distribution of authors by the share of their notes rated as Helpful, Not Helpful, or Needs More Ratings as of 1 July, 2025.

RQ3.c Do the most frequent posters mainly produce helpful notes?

Having established that Community Notes (CN) participants generally do not predominantly produce helpful notes, an important subsequent research question concerns whether this pattern differs among frequent note authors. These frequent contributors, by virtue of their higher activity levels, are often presumed to be more experienced and thus expected to maintain a higher standard of quality in their notes to avoid account restrictions and sustain their participation.

Contrary to this expectation, our analysis reveals that the most active note authors do not predominantly generate helpful notes (Figure 5.9). Among the top 10% of contributors—who collectively produce the majority of notes—only 4.2% (3,117 out of 74,823) surpass the threshold of having at least 30% of their notes rated as helpful. In contrast, among the less frequent authors, who contribute to the remaining 20% of notes, a significantly higher percentage of authors, 8.6% (21,361 out of 202,735), achieve this threshold. This inverse relationship suggests that increased posting frequency may come at the expense of note quality.

These findings challenge the assumption that higher activity correlates with higher expertise or note quality within the CN system. It may indicate that prolific authors prioritize quantity over quality, potentially due to pressures to maintain visibility or avoid account sanctions. Alternatively, this trend could reflect diminishing marginal returns in quality as individual authors produce more content. This dynamic warrants further investigation, particularly concerning the mechanisms that incentivize or discourage quality among highly active contributors, and how platform policies might better support constructive participation.

Figure 5.9. Distribution of authors by the share of Helpful notes, where the Helpfulness level is defined at 30% threshold. The right top area shows the proportion of authors among the top 10% of most active contributors whose 30% or more notes are ranked helpful, while the bottom part displays the proportion of top 10% authors whose helpfulness level is less than 30%. The left hand side demonstrates the same distribution for the remaining (bottom 90%) of authors.

RQ3.d: Which user enrolment statuses are associated with the majority of published notes?

Figure 5.10 illustrates the distribution of Community Notes (CN) participants by user status. The largest share of participants (58.4%) are classified as new users, who have not yet acquired posting privileges and are limited to rating notes authored by others. The second largest group (37.6%) comprises earnedIn users, who possess both the right to post notes and the ability to rate others’ contributions. By contrast, atRisk and locked users—those who have either accumulated low-quality contributions or have had their note-posting privileges revoked—represent a small minority of the population (0.1% and 2.8%, respectively). Given this distribution, it would be reasonable to expect that the majority of notes would be authored by the largest segment of eligible contributors, namely the earnedIn users.

Figure 5.10. The proportion of unique CN users falling into each enrollment type as of July 1, 2025

However, our analysis of note production by enrolment status reveals a counterintuitive pattern (Figure 5.11): atRisk users, despite representing the smallest (0.1%) proportion of the CN population, produce nearly eight times more notes (58 notes on average) than any other contributor group (2-7 notes on average), including the much larger earnedIn segment. This disproportionate output aligns partially with our earlier finding that the most frequent posters tend to produce significantly fewer helpful notes. The elevated activity among atRisk users suggests that prolific note authorship is associated with a higher likelihood of producing content that fails to meet community quality standards, thereby increasing the risk of account restriction.

This dynamic raises important questions about the motivational and behavioural drivers behind such patterns. One possibility is that atRisk users engage in high-volume posting in an attempt to counterbalance prior low-quality ratings, hoping to improve their standing through quantity rather than quality. Alternatively, the behaviour may reflect a persistent misunderstanding of the platform’s quality criteria, or even strategic posting aimed at influencing note visibility regardless of helpfulness outcomes. Such dynamics underscore the need for platform-level interventions that address not only content quality, but also the structural incentives that shape participant behaviour.

Figure 5.11. The average number of notes per user in each enrollment type as of July 1, 2025.

RQ3.e What are the topic and persuasion techniques dynamic across the CN platform?

Figure 5.12 depicts the distribution of the ten most frequently used topics by the top 1,000 Community Notes authors. The most prevalent topics include Politics, Security, and Religion, indicating a strong orientation toward sociopolitical dynamics, moral discourse, and public safety. This thematic focus is consistent with prior work by Borenstein et al. (2025), who—using a different topic classification—also identified Politics as the most prominent domain in both notes citing fact-checking sources and those without such citations. Similarly, our results align with their finding that Healthcare ranks among the top five most discussed themes.

The prominence of these topics offers important insight into the interpretive priorities of highly active CN contributors. The thematic distribution suggests that contributors may be particularly drawn to issues with high public salience and moral significance, where factual accuracy and interpretive framing are both highly contested. In turn, this has implications for the role of Community Notes in shaping the information ecosystem: when annotations cluster around politically or socially charged domains, the fact-checking process itself may become entangled with ideological positioning, value-laden framing, and heightened potential for disagreement. Thus, the concentration of activity within these domains not only reflects contributors’ topical interests but also shapes the epistemic boundaries of the CN platform—potentially amplifying certain narratives while leaving other issue areas comparatively under-annotated.

Figure 5.12. The distribution of topics identified in CNs authored by the top 1000 most active contributors. The topics were identified using the framing classifier by Wu et al., 2023.

Finally, we examined the distribution of persuasion techniques employed by the three principal user categories identified in RQ3.d: (i) users one Not Helpful rating away from losing their note-writing privileges (atRisk status), (ii) active contributors who have earned and retained the right to write Community Notes (earnedIn status), and (iii) users whose note-writing privileges have been revoked (earnedOutAcknowledged status). Our analysis (Figure 5.11) reveals notable differences both in the absence and in the use of persuasion techniques across these groups. Contrary to our initial hypothesis, earnedIn users—those not at risk of losing their privileges—are less likely to produce notes without any persuasion techniques than both atRisk and earnedOutAcknowledged users. Moreover, earnedIn users are significantly more inclined to employ emotional language in their annotations compared to the other two groups.

This finding is counterintuitive, as one might expect experienced and trusted contributors to avoid rhetorical strategies that could be perceived as biased or propagandistic. Instead, it appears that the CN rater community does not strongly penalise such techniques, suggesting that persuasive—even affective—language can coexist with perceptions of helpfulness.

Across all three user categories, four persuasion techniques emerge as consistently prevalent: Loaded Language, Repetition, Name-Calling, and Casting Doubt. Moderate use is also observed for Appeals to Authority figures, Appeals to Fear, Exaggeration–Minimisation, and patriotic framing (Flag-Waving). The remaining 13 identified persuasion techniques are extremely rare, each constituting less than 1% of all notes. Certain strategies—such as Oversimplification, Straw Man, and Appeal to Time—are entirely absent from the dataset.

These results suggest that CN authors, regardless of status, operate within a relatively narrow repertoire of rhetorical strategies, potentially reflecting both the informal norms of the platform and the specific affordances of its short-form annotation structure. The prevalence of emotionally charged and adversarial techniques among even the most trusted contributors raises questions about the role of affect and rhetorical framing in shaping perceived note quality. This finding may also indicate that CN, while positioned as a corrective mechanism for misinformation, functions as much as a site of interpretive contestation—where persuasion and framing are integral to the note-writing process—as it does a purely fact-verification tool.

Figure 5.13. The distribution of persuasion techniques across 3 types of user enrollment statuses. The techniques were inferred using the persuasion technique classifier by Wu et al., 2023.

---

6. Discussion

This study critically evaluates the capacity of the Community Notes (CN) system to meet its intended objectives of generating consensus, delivering timely contextual information, and fostering high-quality, broad-based participation. The results indicate that, rather than consistently converging on agreement, the system frequently produces outcomes characterised by non-agreement or dissensus, both at interim stages and in final note statuses. This tendency calls into question the assumption that algorithmically derived “Diversity of Perspectives” reliably captures genuine collective consensus. A key example of this is the emergence of the “Note Not Needed” (NNN) practice, where contributors employ the CN mechanism to engage in meta-discussion or express disagreement with the existence of a note, rather than to directly assess the veracity of a post. While this behaviour introduces elements of deliberation reminiscent of participatory democratic processes, it diverges from the platform’s stated aim of producing factual, context-rich annotations. The analysis of timeliness reveals a significant limitation: CNs are often published after the peak of a post’s spread. With an average publication delay of nearly four days—and a median of 19.5 hours—many notes are unlikely to meaningfully influence the early stages of misinformation diffusion. This temporal lag is comparable to that of professional fact-checkers, undermining the belief that crowdsourcing necessarily offers a speed advantage.

Patterns of participation reveal further systemic imbalances. A small subset of highly active contributors—27% of authors—produce 80% of all notes, mirroring the Pareto distribution observed in other collaborative platforms and making CNs comparable to professional fack-checkers in terms of the risks of censorship and biases. Additionally, contrary to expectations that frequent contributors would maintain higher note quality, the most prolific authors produce helpful notes at a lower rate than less active peers. This suggests that high posting frequency may compromise quality, potentially due to reduced diligence or strategic behaviour aimed at maximising output rather than accuracy. Moreover, the distribution of user statuses highlights that at-risk users—those close to being locked out—produce disproportionately more notes than active users in good standing, raising questions about the motivations and oversight mechanisms for high-volume contributors.

The content analysis of topics and persuasion techniques reveals that CN authors predominantly focus on politically charged and socially salient issues such as politics, security, religion, and health. Persuasion strategies, including Loaded Language, repetition, name-calling, and casting doubt, are prevalent across all user groups, with emotional framing particularly common among active contributors. This suggests that even among notes rated as helpful, rhetorical devices with propagandistic potential remain widespread, indicating that the CN rater community may not strongly penalise emotionally charged or biased language. This tolerance could undermine the platform’s goal of fostering unbiased, evidence-based contributions.

7. Conclusions

This research highlights fundamental tensions in the Community Notes system between its stated purpose and its operational realities. While designed to foster consensus through algorithmic diversity, the system often yields dissensus, with many notes remaining unrated or marked as unnecessary. The use of NNN as a vehicle for intra-community debate further underscores the gap between intended and actual use, suggesting a latent demand for discussion spaces that CN currently does not formally accommodate.

Timeliness remains a critical bottleneck. Despite policy changes that have accelerated the note publication process, the CN system still lags behind the viral spread of online content, particularly misinformation. The effectiveness of CNs in mitigating harm is therefore constrained by both procedural delays and the broader dynamics of information diffusion on social media platforms.

Participation dynamics reflect an over-reliance on a small, hyperactive subset of contributors, whose output is not necessarily of higher quality than that of less active peers. This concentration of authorship parallels criticisms historically levelled at expert-driven fact-checking, namely the risks of bias and representational imbalance. The finding that at-risk users contribute disproportionately to note creation raises further concerns about whether volume incentives overshadow quality control.

Finally, the persistence of persuasive and emotionally charged rhetoric—even in notes rated as helpful—points to limitations in the current evaluation framework. If the rater community does not systematically penalise biased framing, CNs may inadvertently reproduce the same affective and polarising discourse they are intended to counteract. Future research should explore mechanisms for integrating deliberative, open-discussion features without compromising the system’s informational objectives, as well as revising quality metrics to account for rhetorical neutrality alongside factual accuracy.

---

8. References

Augenstein, I., Bakker, M., Chakraborty, T., Corney, D., Ferrara, E., Gurevych, I., Hale, S., Hovy, E., Ji, H., Larraz, I. and Menczer, F., 2025. Community Moderation and the New Epistemology of Fact Checking on Social Media. arXiv preprint arXiv:2505.20067.

Barnett, K. (2025). AI Chatbots Can Now Write Community Notes on X, AdWeek, July, https://www.adweek.com/media/exclusive-ai-chatbots-can-now-write-community-notes-on-x/.

Bassett, K. (2025). X’s Community Notes and the South Asian Misinformation Crisis. Center for the Study of Organized Hate. https://www.csohate.org/2025/06/30/x-community-notes-south-asia/.

Borenstein, N., Warren, G., Elliott, D. and Augenstein, I., 2025. Can Community Notes Replace Professional Fact-Checkers?. arXiv preprint arXiv:2502.14132.

Capraro, V., & Vanzo, A. (2019). The power of moral words: Loaded language generates framing effects in the extreme dictator game. Judgment and Decision Making. https://doi.org/10.1017/S1930297500004356.

Chuai, Y., Pilarski, M., Renault, T., Restrepo-Amariles, D., Troussel-Clément, A., Lenzini, G., & Pröllochs, N. (2024). Community-based fact-checking reduces the spread of misleading posts on social media. arXiv preprint arXiv:2409.08781.

Chuai, Y., Sergeeva, A., Lenzini, G., & Pröllochs, N. (2024). Community Fact-Checks Trigger Moral Outrage in Replies to Misleading Posts on Social Media. arXiv preprint arXiv:2409.08829.

Chuai, Y., Tian, H., Pröllochs, N., & Lenzini, G. (2024). Did the Roll-Out of Community Notes Reduce Engagement With Misinformation on X/Twitter?. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW2), 1-52.

Crisan, A. and McNutt, A.M., 2025, April. Linting is People! Exploring the Potential of Human Computation as a Sociotechnical Linter of Data Visualizations. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (pp. 1-10).

Drolsbach, C. P., Solovev, K. & Pröllochs, N. (2024) Community notes increase trust in fact-checking on social media, PNAS Nexus, 3(7), https://doi.org/10.1093/pnasnexus/pgae217

Flamini, D., 2019. Most Republicans don’t trust fact-checkers, and most Americans don’t trust the media. Poynter Institute. Retrieved May, 12, p.2022.

Greenwood, P.E. and Nikulin, M.S., 1996. A guide to chi-squared testing. John Wiley & Sons.

Kangur, U., Chakraborty, R., & Sharma, R. (2024). Who Checks the Checkers? Exploring Source Credibility in Twitter's Community Notes. arXiv preprint arXiv:2406.12444.

de Keulenaar, E. (2025) From Twitter to X: demotion, community notes and the apparent shift from adjudication to consensus-building, SSRN preprint, 10.2139/ssrn.5165083

Kovermann, E., (2025). Community Notes as Algorithmic Consensus-Making A Critical Review, Master’s Thesis, University of Amsterdam.

Lewandowsky S, Ecker UK, Seifert CM, Schwarz N, Cook J. Misinformation and Its Correction: Continued Influence and Successful Debiasing. Psychol Sci Public Interest. 2012 Dec;13(3):106-31. doi: 10.1177/1529100612451018. PMID: 26173286.

Martel, C., Allen, J., Pennycook, G. and Rand, D.G., 2024. Crowds can effectively identify misinformation at scale. Perspectives on Psychological Science, 19(2), pp.477-488.

Matamoros-Fernández, A., & Jude, N. (2025). The importance of centering harm in data infrastructures for ‘soft moderation’: X’s Community Notes as a case study. New Media & Society, 27(4), 1986-2011. https://doi.org/10.1177/14614448251314399 (Original work published 2025)

Murphy, J. (2025). Use of X’s Community Notes has plummeted in 2025, data shows, NBC News, 6 June, https://www.nbcnews.com/tech/social-media/x-twitter-community-notes-disappear-data-rcna210710.

Nenno, S. (2025, February 20). Do Community Notes Have a Party Preference? – Digital Society Blog. Alexander von Humboldt Institute for Internet and Society gGmbH. https://www.hiig.de/en/community-notes-analysis/.

Newton, C. (2025) Meta surrenders to the right on speech, Platformer, 7 January, https://www.platformer.news/meta-fact-checking-free-speech-surrender/.

Pahwa, N. (2023). What I Saw in Elon Musk’s Truth Army, Slate, https://slate.com/technology/2023/07/twitter-community-notes-elon-musk-fact-checking.html

Powell, T. and Sammut-Bonnici, T., 2014. Pareto analysis.

Renault, T., Amariles, D. R., & Troussel, A. (2024). Collaboratively adding context to social media posts reduces the sharing of false news. arXiv preprint arXiv:2404.02803.

Stencel, M., Ryan, E., & Luther, J. (2024). With half the planet going to the polls in 2024, fact-checking sputters. Duke Reporters’ Lab.

Tagliabue, C., Assecondi, S., Cristoforetti, G., & Mazza, V. (2020). Learning by task repetition enhances object individuation and memorization in the elderly. Scientific Reports, 10. https://doi.org/10.1038/s41598-020-75297-x.

Vraga, E. K., & Bode, L. (2017). Using Expert Sources to Correct Health Misinformation in Social Media. Science Communication, 39(5), 621-645. https://doi.org/10.1177/1075547017731776 (Original work published 2017)

Wirtschafter, V. and Majumder, S., 2023. Future challenges for online, crowdsourced content moderation: evidence from Twitter’s community notes. Journal of Online Trust and Safety, 2(1).

Wojcik, S., Hilgard, S., Judd, N., Mocanu, D., Ragain, S., Hunzaker, M. B., ... & Baxter, J. (2022). Birdwatch: Crowd wisdom and bridging algorithms can inform understanding and reduce the spread of misinformation. arXiv preprint arXiv:2210.15723.

Wu, B., Razuvayevskaya, O., Heppell, F., Leite, J.A., Scarton, C., Bontcheva, K. and Song, X., 2023, July. SheffieldVeraAI at SemEval -2023 Task 3: Mono and Multilingual Approaches for News Genre, Topic and Persuasion Technique Classification. In Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval -2023) (pp. 1995-2008).

X, Community Notes. (n.d.). Download data, https://communitynotes.x.com/guide/en/under-the-hood/download-data