Community Notes — Evidence of Their Effectiveness Against Disinformation

On March 18, 2025, Meta began testing its “Community Notes” in the United States, a feature that allows selected users to add context to posts that might be misleading on Facebook, Instagram and Threads. This system replaces third party fact checking and assigns users the responsibility of flagging mistakes or inaccuracies in posts. TikTok also announced a similar program called Footnotes, although as a complement to the work of fact-checkers.

Community Notes are a collaborative tool that Twitter (now X) has been testing since 2021, based on the concept of “Wisdom of Crowds,” which suggests that large groups of people can make collective judgments that are as accurate as those of individual experts (see here, for example)

How Do They Work?

---

According to Meta, and following X’s model, in order to write notes, people must first sign up and meet certain criteria, like living in the United States, having a verified phone number and not having violated their policies against severe harms (such as terrorism, child sexual exploitation, and fraud and scams).

Selected contributors can write notes for posts they deem misleading. Each note must include background information, an insight or point of view and at least one link to a source that supports the note. Once the note is submitted, other contributors rate it and it only gets published if enough contributors agree that the note is helpful.

The key aspect is that the rating is not determined by a majority, but through consensus among contributors who normally disagree with each other on the helpfulness of a note. The reasoning is that if people who don’t usually agree with each other consider that a note is helpful, it probably is for people with different points of view.

Are They Effective at Tackling the Disinformation Flow?

---

Hasta el momento, Meta, y a diferencia de X, solo ha publicado datos generales: aproximadamente 70 mil usuarios han escrito 15 mil notas, con tasa de publicación del 6%. La evidencia disponible proviene principalmente de la experiencia en X.

Quality of Information in Community Notes

---

The quality of Community Notes can be determined based on the sources used to support the note. A study examined over 500,000 notes on X posted between January 2021 and January 2024, and observed that among the top 50 quoted sources there are renowned media outlets (CNN, Reuters, BBC, The Guardian, The New York Times), trustworthy institutions such as the National Institutes of Health (NIH) and the World Health Organization (WHO) and fact-checking organizations.

Another study, conducted by the Spanish fact-checking organization Maldita, analyzed over a million notes written in 2024 and found that approximately 3.7% included links to certified fact-checkers. On a global scale, this puts fact-checkers as the third most quoted source, behind X and Wikipedia. Even in other languages (German, French, Italian, Spanish, Polish), there is always at least one fact-checker among the top 20 most quoted sources.

A smaller study evaluated 650 notes addressing COVID-19 vaccination posted between December 2022 and December 2023, and found that 97% were supported by scientific evidence, while only a minor fraction was biased or completely inaccurate. Regarding sources, almost half quoted highly credible primary sources (scientific journals or official websites), 44% quoted known secondary sources (media outlets or fact-checkers), and only 7% quoted low-quality sources.

However, the system is not without vulnerabilities. Because it is collaborative, it can be manipulated. During the pilot phase of the program (Birdwatch), a study concluded that for certain topics, like the 2020 U.S. election or COVID-19, a large number of users rated trustworthy sources quoted by other people as low-quality. According to the authors, this shows that “there is a group of people that is trying to deceive the system to serve their common interest.” Since the publication of a note depends on reaching consensus between evaluators, these practices can reduce the likelihood that well-founded corrections will become visible.

The Barrier of Consensus

---

For a note to be visible, X requires consensus between people with different points of view. In theory, this would guarantee diversity and balance. In practice, consensus works as a filter that often delays or blocks the publication of well-founded notes.

A study of the Center for Countering Digital Hate (CCDH) analyzed notes proposed between March and August 2024 about the U.S. election, and showed important limitations. Although many notes were accurate and quoted quality sources, 74% were never shown to users for not being rated “helpful.” As a result, posts with incorrect information about the election had over 2.2 billion views without their corresponding correction. Among them were posts with false claims about Donald Trump, or posts that claimed that Democrats were “importing illegal voters,” that voting systems were unreliable or that the 2020 presidential election had been rigged. Even when the posts included Community Notes, tweets with disinformation received 13 times more views than their Community Notes.

A more comprehensive analysis from the Digital Democracy Institute of the Americas (DDIA) examined over 1.7 million notes posted between 2021 and 2025 in 55 languages and also exposed relevant limitations in the effectiveness of the program. Although the number of users that submit notes grew continuously and doubled in 2024, only a small fraction is shown: 7.1% of all submissions in English and only 4.7% in Spanish. In addition, another important fraction of all contributions does not even go through the consensus process because they are never evaluated in the first place: over 17% of English notes and 15% of Spanish notes remain unevaluated, which leaves most potential corrections out of circulation. The time it takes for a note to go from submission to publication has improved — from more than 100 days in 2022 to an average of 14 days in 2025 — but the delay is still excessive to counteract the speed at which disinformation expands online.

The report from Maldita also revealed that only 8.3% of the proposed notes becomes visible to users; i.e., nine out of ten corrections are never shown. However, notes that quote verifications from professional fact-checkers have better chances: 12% becomes visible. In addition, they get posted more quickly. After the tweet is posted, these notes are proposed more quickly and, once they are evaluated, reach consensus in a shorter amount of time, which makes them especially valuable against the speed at which disinformation spreads.

These results open the debate about the logic behind consensus, which does not always guarantee that the best contributions will be seen, and show the need for rethinking a system that values more robust sources and expert knowledge.

Can They Stop the Spread of Disinformation?

---

The main question is whether a note that flags or puts into context a disinformation tweet can stop its spread. In summary, do they persuade users not to share these posts flagged as fake?

A study published in September 2025 found that when a note reaches the required consensus and becomes visible, that tweet has fewer shares, spreads less and loses traction. The study analyzed over 40,000 posts on X for which notes had been proposed between March and June 2023. To estimate the effect of published notes, the authors compared posts with visible notes with posts for which a note was proposed, but not attached, since it hadn’t reached the required consensus. They made sure to compare posts that, before the note was published, would have had the same performance in terms of virality.

Results show that in the 48 hours after the community note was attached, posts with a visible correction had reductions of 46% in reposts, 44% in likes, 22% in replies and 14% in views, compared to posts without visible notes. Considering the total life span of each post, the impact is still significant: 11.6% fewer reposts, 13.3% fewer likes, 6.9% fewer replies, and 5.5% fewer views. Community Notes not only slow down the initial diffusion, they also affect how far and how widely misleading content can spread.

To understand the real reach of each tweet and capture its total virality, the study conducted an in-depth and original analysis known as “diffusion cascades” using different metrics. Breadth measures how many users repost directly from the original post. It represents a post’s initial impact and usually depends on the number of followers, if the post is attention-grabbing, etc. Depth, on the other hand, measures how many times that content is shared through different levels of users. Someone sees it and shares it, then someone else sees that repost and shares it again, and so on. The longer the chain, the deeper the diffusion. Lastly, structural virality combines both dimensions — breadth and depth — to measure the “nods” in the content’s diffusion. This type of analysis doesn’t only take into consideration how many people shared a post, but how that network was built; whether it was a centralized explosion (many direct reposts) or a more complex spread through the network.

Considering this broader analysis, results show that, although, on average, notes reduce the depth and structural virality of posts with disinformation, the effect on breadth is smaller. This suggests that the system works better at breaking the chain-based spread, but it doesn’t affect the number of people who repost it at the beginning.

This difference is important because a post doesn’t need millions of likes to be dangerous; if it manages to enter a network and spread through reposting chains, it can continue circulating for days. Community Notes seem to act on this capacity to remain active over time through chains of users.

The study also identified conditions where notes work better. Speed is key: the quicker the note is posted, the less viral it gets. Notes that appear late have almost no effect. The type of content also matters: corrections have a bigger impact on manipulated images or videos than on posts that only contain text. Lastly, writing matters: notes that are relatively long, written in simple language, are more likely to perform better.

Do They Replace or Complement Fact-Checkers?

---

When Mark Zuckerberg announced the launch of Community Notes, he explained that the new program would “get rid” of fact-checkers. However, evidence shows differently. Community Notes don’t replace fact-checkers, they complement them.

Data shows that when a Note is supported by a professional fact-check, it builds more trust among users and is more likely to overcome the consensus barrier, thus becoming visible more quickly and more often. Something similar happens when the disinformation claim is complex, either because it requires several steps of reasoning or because it combines data from different sources. In those cases, notes tend to quote professional fact-checks more often. Since one of the limitations of this collaborative system is making notes visible, using verifications from fact-checkers becomes a key resource.

From the point of view of fact-checking organizations, it’s not about competition, but cooperation. As Maldita proposes, Community Notes and fact-checkers can (and should) work together, combining the speed and scale of the collaborative model with the methodological rigor of professional fact-checking. Fact-checkers can carry out in-depth research that is out of reach for amateur users, while Community Notes publicize their work.

In conclusion, the argument is not whether one replaces the other, but how they could better complement each other within the same ecosystem to stop disinformation.

Literature Consulted

---

1. Allen, J., Arechar, A. A., Pennycook, G., & Rand, D. G. (2021). Scaling up fact-checking using the wisdom of crowds. Science Advances, 7(36), eabf4393. https://doi.org/10.1126/sciadv.abf4393
2. Allen, M. R., Desai, N., Namazi, A., Leas, E., Dredze, M., Smith, D. M., & Ayers, J. W. (2024). Characteristics of X (formerly Twitter) Community Notes addressing COVID-19 vaccine misinformation. JAMA, 331(19), 1670–1672. https://doi.org/10.1001/jama.2024.4800
3. Borenstein, N., Warren, G., Elliott, D., & Augenstein, I. (2025). Can Community Notes replace professional fact-checkers? In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) (pp. 535–552). https://doi.org/10.18653/v1/2025.acl-short.42
4. Braga, R., Tardáguila, C., & Soares, M. (2025, July 9). A deep dive into X’s Community Notes: An analysis of English and Spanish contributions between 2021 and 2025. The Digital Democracy Institute of the Americas. https://ddia.org/en/a-deep-dive-into-xs-community-notes-report
5. Center for Countering Digital Hate. (2024, October 30). Rated not helpful: How X’s Community Notes system falls short on election disinformation [Report]. https://counterhate.com/wp-content/uploads/2024/10/CCDH.CommunityNotes.FINAL-30.10.pdf
6. Fundación Maldita. (2025). Faster, trusted, and more useful: The impact of fact-checkers in X’s Community Notes [Report]. https://files.maldita.es/maldita/uploads/2025/02/maldita_informe_community_notes_2024.pdf
7. Fundación Maldita. (2025, January 17). Por qué las Notas de la Comunidad y los verificadores pueden (y deben) trabajar juntos, y cómo hacerlo para que funcione. https://maldita.es/nosotros/20250117/notas-comunidad-verificadores-juntos-funcione/
8. Kangur, U., Chakraborty, R., & Sharma, R. (2024). Who checks the checkers? Exploring source credibility in Twitter’s Community Notes. arXiv. https://doi.org/10.48550/arXiv.2406.12444
9. Saeed, M., Traub, N., Nicolas, M., Demartini, G., & Papotti, P. (2022). Crowdsourced fact-checking at Twitter: How does the crowd compare with experts? arXiv. https://doi.org/10.48550/arXiv.2208.09214
10. Slaughter, I., Peytavin, A., Ugander, J., & Saveski, M. (2025). Community notes reduce engagement with and diffusion of false information online. Proceedings of the National Academy of Sciences of the United States of America, 122(38), e2503413122. https://doi.org/10.1073/pnas.2503413122