Community Notes is a crowdsourced system for adding contextual information to posts on social media, best known for its implementation on X, formerly Twitter. Eligible contributors write short notes and assess notes written by others. A note becomes eligible for prominent display when the system determines that contributors with differing rating histories agree that it is helpful. The distinctive feature is therefore not simply public participation, but an attempt to identify agreement across disagreements rather than count favorable votes alone.
Twitter introduced the system as Birdwatch in January 2021 and renamed it Community Notes in November 2022. It developed into a prominent example of participatory fact-checking, combining volunteer judgment with statistical ranking. Its code and downloadable contribution data have also made it an important subject for research into online information quality and platform governance. Related systems have subsequently been developed or announced by other platforms.
Community Notes does not function as a comprehensive verification service, and the absence of a note does not establish that a post is accurate. Its operation depends on which posts attract contributors, the evidence available to them, and whether sufficient agreement develops before attention moves elsewhere. Understanding the system requires distinguishing the accuracy of individual notes, the rules that determine their visibility, and their effects on readers and information circulation.
Origins and development
Community Notes emerged from efforts to address misleading information without assigning every disputed claim to a platform employee or a professional fact-checking organization. Before Birdwatch, Twitter already used several forms of intervention, including labels, links to authoritative information, and restrictions on the circulation of particular posts. Birdwatch introduced a different institutional arrangement: users could supply contextual explanations, while a rating system would decide which contributions merited broader visibility.
Twitter announced the Birdwatch pilot on 25 January 2021. Initially, participation was restricted to a relatively small group in the United States, and notes were primarily encountered through a separate Birdwatch interface. This arrangement allowed the company to test the writing and rating processes before displaying notes widely alongside tweets. The project’s published rationale emphasized the possibility that people with different perspectives could recognize a useful explanation even when they disagreed about the underlying subject.
During 2022, Twitter expanded participation and the visibility of notes. In October, it made qualifying notes visible to users across the United States. Following Elon Musk’s acquisition of Twitter, the service was renamed Community Notes in November 2022. International expansion followed, with additional countries and languages admitted over time. These changes transformed the project from a geographically limited pilot into a feature used across a multinational platform.
Its technical architecture also expanded. Notes originally addressed individual text posts, but later features allowed context to accompany images and videos and to appear on other posts containing matching media. This mattered because misleading visual material often circulates through repeated uploads rather than a single widely shared original. An old disaster photograph, for example, can be repeatedly presented as evidence of a new event, making post-by-post annotation inefficient.
The development of Community Notes was not a single completed transition from experiment to fixed product. Eligibility requirements, scoring rules, contributor incentives, interfaces, and distribution mechanisms have changed. Historical comparisons therefore require attention to the version of the system being studied. Results obtained from a small, English-language Birdwatch pilot cannot automatically be applied to a larger multilingual service with different contributors and visibility rules.
Participation, note writing, and contributor roles
Community Notes separates ordinary platform use from participation in the annotation system. Reading a displayed note does not require contributor status. Writing and rating notes, however, requires enrollment and satisfaction of eligibility conditions. X has used criteria involving account age, recent rule violations, and a verified telephone number associated with a trusted carrier. These requirements are intended to increase the cost of creating disposable contributor accounts, although they do not establish expertise or guarantee independent behavior.
Contributors participate under aliases distinct from their ordinary account identities. This separation is intended to reduce reputational pressures and make it more difficult to reward or punish a contributor through their public profile. It is not equivalent to complete anonymity: the platform retains account information, and the content or timing of contributions can sometimes reveal interests or affiliations. The arrangement illustrates the tension between accountability and data privacy in systems that ask volunteers to evaluate contentious material.
A contributor can propose a note explaining why a post may be misleading or why additional context is relevant. Useful notes typically isolate a checkable issue, give a concise explanation, and cite evidence. A note about a photograph might link to its earlier publication; one about a numerical claim might identify the original dataset and explain a missing denominator. The objective is to supply information that changes how the post should be understood, not merely to register disagreement.
Other contributors assess notes using helpfulness judgments and structured reasons. Relevant considerations include whether the note is supported by reliable sources, directly addresses the claim, uses clear language, and avoids argumentative or inflammatory wording. Contributors can also identify defects such as irrelevant evidence, unsupported assertions, or a misrepresentation of the original post. These judgments provide both the inputs for scoring and feedback about recurring weaknesses in note writing.
Contributor privileges and reputation are connected to previous performance. The system distinguishes a contributor’s record of writing notes that become helpful from their record of rating notes in ways that align with later outcomes. Rating performance can affect access to writing privileges. This creates an incentive to anticipate judgments that will survive broader review, rather than simply support notes that express a preferred position. It also makes the design of the reputation system consequential: an incorrect consensus can affect both a note’s visibility and the standing of people who assessed it.
The cross-perspective scoring mechanism
The central technical problem is to distinguish broadly useful context from coordinated approval. A simple vote count can reward the largest or most active group, particularly when a claim concerns a divisive subject. Community Notes instead uses a “bridging” approach: a note should receive support from contributors who have not consistently agreed with one another in the past. The intended signal is agreement that crosses observed patterns of disagreement.
An important component of the published scoring approach is matrix factorization, a technique also used in recommendation systems. Ratings can be represented as a partially observed table, with contributors in one dimension and notes in the other. Statistical models estimate patterns in that table, including whether some contributors tend to rate particular kinds of notes similarly. The resulting dimensions are inferred from behavior; they are not direct declarations of political identity, nationality, or subject expertise.
A simplified version of the modeling idea can be written as
where is the predicted rating by contributor of note , is an overall baseline, represents a contributor-specific tendency, represents a note-specific component, and the interaction between and captures patterned disagreement. This equation illustrates the modeling principle rather than reproducing every rule in the production system. Regularization and additional constraints help prevent sparse or unusual rating patterns from dominating estimates.
The practical aim is to identify notes with a strong helpfulness signal after accounting for differences among raters. Agreement from contributors who usually rate similarly supplies less evidence of bridging than agreement spanning different rating patterns. Conversely, a note attracting many positive ratings from one cluster may fail to become helpful if other relevant contributors consistently reject it. This differs from ordinary majority rule: accumulating votes is not necessarily sufficient.
The operational system includes more than a single fitted model. Published implementations have used multiple scorers, eligibility rules for ratings, thresholds, checks on unusual patterns, and mechanisms governing changes of status. A note can remain unresolved because evidence of agreement is insufficient, rather than because the system has established that it is wrong. The distinction is fundamental: the algorithm estimates a form of cross-perspective helpfulness, not truth itself. Its outputs remain dependent on contributor coverage, available evidence, and the assumptions encoded in its rules.
Evidence, sourcing, and the meaning of a helpful note
Community Notes occupies a space between an annotation service and a fact-checking organization. Its notes often perform fact-checking functions, but the system does not require every intervention to pronounce an entire post true or false. Context can be useful when a post is literally accurate yet omits a crucial qualification. A graph may use authentic figures while concealing a change in measurement practice; a quotation may be genuine but shortened in a way that reverses its implication.
Source quality is therefore central. A link is not sufficient merely because it exists. The cited material must support the specific correction being made, and contributors must distinguish what a source directly establishes from what they infer. Original records, official datasets, complete transcripts, and contemporaneous documents can resolve some questions more directly than secondary commentary. In other cases, specialist reporting or scholarly analysis is needed to interpret evidence that would otherwise be easy to misunderstand.
The relationship with journalism is complementary but not identical. Journalists may conduct interviews, inspect documents that are not public, or undertake lengthy investigations. Community Notes usually operates with short explanations and evidence already accessible online. A note may draw on reporting without replicating the reporting process. Its usefulness can consequently depend on an external information infrastructure maintained by news organizations, public institutions, researchers, and archivists.
Some categories of claim are especially amenable to concise annotation. Misattributed quotations, outdated images, incorrect dates, and numerical comparisons with missing context often have relatively determinate resolutions. Predictions, satire, value judgments, and ambiguous rhetoric are harder to handle. A contributor may mistake disagreement about priorities for a factual error, or may write a factually correct note that does not actually contradict the post. Assessing relevance is thus as important as checking isolated statements.
A helpful status should not be understood as a universal endorsement of every sentence or source in a note. It records the outcome of a particular assessment process at a particular time. Sources can be revised, new evidence can emerge, and a note’s status can change. Similarly, a post with no visible note may never have been reviewed, may have attracted an unresolved draft, or may concern a claim that contributors could not assess. The interface communicates a selective intervention, not a complete inventory of verified and unverified content.
Display, distribution, and platform enforcement
The visibility of a note is a separate stage from its creation. Contributors may be able to inspect and rate a proposed note before ordinary users see it prominently beneath a post. Once the note satisfies the relevant criteria, it can appear in the main interface with its explanatory text and supporting links. This staged process is intended to keep unreviewed annotations from acquiring the apparent authority of an accepted correction.
The distinction between note creation and public display has important consequences. A widely circulating post may accumulate substantial attention while a proposed note is still being assessed. Even after a note becomes visible, not every person who previously encountered the post will see the correction. X has used notifications and other interface mechanisms to bring notes to the attention of users who interacted with a post, but such mechanisms cannot guarantee that the original audience receives or reads the added context.
Media-level notes extend the system beyond the individual post. If the same misleading image is uploaded by many accounts, attaching context to matching copies can reduce duplicated work. This is especially useful for recycled photographs, edited screenshots, and videos presented with a false date or location. Matching nevertheless creates a relevance problem: the same image can appear in a misleading assertion, a legitimate news report, or a post explicitly debunking it. The relationship between the image and its accompanying claim still matters.
Community Notes is also distinct from content moderation that removes material or limits account access. A note generally adds information while leaving the underlying post available. Other platform rules can apply independently, including rules concerning illegal material, harassment, or manipulated media. A post may therefore receive a note without violating a removal policy, or violate a policy regardless of whether contributors have written a note.
This distinction is relevant to freedom of expression, but it does not make annotation consequence-free. A prominently displayed correction can affect reputation, engagement, advertising eligibility, or an audience’s willingness to trust an account. The platform still controls the interface, eligibility rules, ranking machinery, and links between note status and other product features. Community participation redistributes part of the judgment process; it does not eliminate the platform’s institutional authority.
Research on accuracy, persuasion, and circulation
Research on Community Notes asks several different questions that require different kinds of evidence. One concerns whether accepted notes are accurate. Another asks whether exposure changes readers’ beliefs. A third examines whether notes reduce sharing or subsequent engagement. These outcomes are related but not interchangeable: an accurate correction may arrive too late to affect circulation, while reduced engagement does not by itself establish improved understanding.
An early experimental study by Jennifer Allen, Cameron Martel, and David Rand, published in 2022 in the Proceedings of the National Academy of Sciences, examined crowdsourced fact-checking through Birdwatch. It found that notes identified as helpful could reduce belief in misleading claims under experimental exposure. This provided evidence that selected crowd-written explanations could influence readers, rather than merely express contributor preferences. The study did not establish that every note was accurate or that the operational system reached all relevant users.
The distinction between experimental efficacy and real-world effectiveness is essential. In an experiment, participants can be assigned to see a post with or without a note, supporting stronger conclusions about causation. On a platform, exposure depends on feeds, user choices, timing, and recommendation systems. People who encounter notes may differ systematically from those who do not. Observational comparisons therefore require careful controls and explicit assumptions.
Studies of platform data have examined changes in reposting, liking, and other engagement around the moment notes become visible. Some report reduced engagement after attachment, while other analyses emphasize that much of a post’s circulation may already have occurred. These findings need not conflict. A note can reduce the rate of new sharing after publication while having only a limited effect on the total audience accumulated before it appeared.
Coverage and selection create further challenges. Notes that achieve helpful status are not a random sample of misleading posts. They are selected first by contributor attention and then by the scoring process. Evaluating only published notes can therefore produce an optimistic picture of the overall system if large amounts of misleading content remain unannotated. Conversely, counting every unresolved draft as a failed correction would ignore drafts that were themselves inaccurate or irrelevant. A rigorous assessment separates accuracy, coverage, speed, exposure, and behavioral effects rather than compressing them into a single success rate.
Coordination, bias, and unresolved limitations
Community Notes is designed partly in response to motivated reasoning: people may scrutinize unwelcome claims more aggressively than claims that support their existing views. Cross-perspective agreement can reduce some consequences of this tendency, but it cannot eliminate selective attention. Contributors still decide which posts to inspect, which sources to consult, and which notes to rate. Topics that attract organized interest may receive extensive scrutiny while less prominent errors remain untouched.
Coordinated participation presents a related problem. A group can direct members to write notes, approve favorable explanations, or reject unfavorable ones. Because the system does not rely on raw vote totals alone, coordination does not automatically produce a helpful status. Nevertheless, sophisticated coordination can seek to influence rating histories, contributor reputation, or the availability of disagreement-spanning support. Eligibility checks and statistical safeguards raise obstacles without proving that every contribution is independent.
The bridging requirement creates a trade-off between acceptance and coverage. On subjects where contributors agree across their usual divisions, it can produce corrections with unusual credibility. On deeply polarized questions, accurate notes may fail to secure sufficient cross-perspective support. Conversely, contributors from different clusters may share the same mistaken assumption. Agreement is evidence about the distribution of judgments, not a logical guarantee that those judgments correspond to reality.
Participation is also uneven across languages, regions, and subject areas. Access depends on platform eligibility, time, familiarity with the interface, and the availability of usable sources. The digital divide therefore affects more than who reads online information; it can shape who has the capacity to annotate it. Small language communities or specialist topics may lack enough active contributors to assess notes quickly, even when competent evidence exists.
Contributor welfare and expertise raise additional questions. Volunteers may encounter disturbing material, persistent conflict, or efforts to infer their identities. Specialists may find that a technically correct explanation is difficult to compress into a short note or to communicate to non-specialist raters. These limitations are not unique to Community Notes, but they become important when a platform relies on volunteer labor for high-stakes information. Evaluating the system requires attention to the conditions under which participation is sustained, not only the notes that eventually appear.
Transparency, public data, and institutional accountability
A distinctive feature of Community Notes is the publication of scoring code, documentation, and downloadable data about notes, ratings, and status histories. This enables outside researchers to examine parts of the system in greater detail than is possible for many platform interventions. Public materials can support replication of scoring procedures, analysis of contributor behavior, and investigation of the time between note creation and display.
Open code does not, however, make the entire platform transparent. The surrounding recommendation system determines which posts gain attention and which users encounter a correction. Researchers may lack complete exposure data, information about deleted posts, or access to all measures used to protect the system from abuse. Reproducing a note score is different from reconstructing the full sequence by which a post and its annotation reached particular audiences.
The data also require careful interpretation. A note’s creation time, its first helpful status, subsequent status changes, and its display history represent different events. A dataset downloaded on one date may not capture later deletions or revisions. Some measurements concern posts, others notes, contributors, or ratings. Treating these units as interchangeable can lead to misleading conclusions, such as inferring audience-wide exposure from the existence of a publicly available note.
Transparency creates privacy and adversarial risks alongside research benefits. Detailed behavioral records can help identify coordination, but they can also allow observers to infer patterns about contributors. Publishing precise scoring rules makes independent scrutiny possible while potentially helping strategic participants search for weaknesses. This is a familiar tension in cybersecurity and institutional auditing: secrecy can conceal defects, whereas disclosure can reveal opportunities for manipulation.
The relationship to public participation is therefore qualified. Contributors have a substantive role in proposing and assessing context, but they do not collectively control every rule governing the system. The platform determines admission requirements, product priorities, and software deployment. Community Notes resembles neither a conventional editorial office nor a self-governing public institution. Its accountability rests on a combination of published procedures, external research, contributor practices, and decisions made by the platform operator.
Adoption beyond X and related approaches
Community Notes belongs to a broader history of collaborative knowledge production. Wikipedia demonstrates how volunteers can build and revise a large reference work, but its procedures differ substantially. Wikipedia organizes work around articles, sources, editing histories, and editorial discussion. Community Notes organizes work around short annotations attached to particular posts and uses a statistical process to decide which annotations become widely visible. Participation alone does not make the two systems institutionally equivalent.
Professional fact-checking provides another comparison. Fact-checking organizations generally select claims, investigate them, and publish under identifiable editorial responsibility. Community Notes distributes selection and assessment among contributors and can operate at a scale difficult for a single newsroom to match. It also depends heavily on evidence produced elsewhere. In practice, crowd annotation and professional verification can be mutually supporting rather than exclusive alternatives: reporting supplies evidence, while annotations place that evidence next to circulating claims.
Other platforms have experimented with similar systems. YouTube announced a test in June 2024 allowing contributors to propose notes that supplied context to videos. In January 2025, Meta announced that it would replace its third-party fact-checking program in the United States with a community-notes approach on Facebook, Instagram, and Threads. These initiatives extended interest in cross-perspective annotation beyond the platform where Birdwatch originated. Their policies, interfaces, and relationships to enforcement must nevertheless be assessed separately rather than treated as copies of one fixed model.
Artificial intelligence introduces another possible role division. Large language models can assist with finding candidate claims, locating sources, drafting explanations, or translating notes. Such uses do not remove the need to establish whether a source actually supports a correction. A fluent generated explanation may contain an unsupported inference or a nonexistent citation, while multiple systems can repeat the same error. Human ratings and source verification address different problems from text generation.
The technical components are consequently more portable than the institutional results. A platform can adopt a bridging algorithm without reproducing X’s contributor population, participation incentives, public data practices, or distribution mechanisms. Video services, messaging systems, and public social networks also differ in how quickly content spreads and where a note can be displayed. Comparative research must examine these conditions directly: agreement among raters is only one stage in the larger process by which contextual information is produced, selected, and encountered.