Community Notes: Crowdsourced Moderation

• Community Notes is a crowdsourced fact-checking and content moderation system that relies on diverse user contributions to add context to potentially misleading posts.
• It employs a bridging-based rating algorithm to ensure that only notes with cross-perspective support are published, mitigating partisan bias.
• The system integrates human expertise with AI-augmented workflows, addressing scalability, timeliness, and participation inequality challenges.

Community Notes is X’s crowd-sourced fact-checking and content-moderation system, launched in January 2021 as Birdwatch and later renamed after the platform’s shift from Twitter to X. In its basic form, contributors write short contextual notes on posts they consider misleading, other contributors rate those notes, and only notes that obtain sufficient support under a bridging-based algorithm are shown publicly beneath the original post. The research literature treats Community Notes as the first large-scale deployment of crowd-based contextual moderation and studies it simultaneously as a moderation mechanism, a socio-technical consensus system, a data resource, and an increasingly hybrid human–AI workflow (Mohammadi et al., 10 Oct 2025, Slaughter et al., 18 Feb 2025).

1. Emergence and institutional scope

Community Notes originated as Birdwatch on Twitter/X in January 2021 and now operates as Community Notes on X. The system is described as a large-scale crowdsourced content-moderation initiative in which vetted users, called Contributors, add context to posts they think are misleading. The workflow is not limited to note writing: contributors also rate notes as helpful, somewhat helpful, or unhelpful, and the system uses those ratings to determine whether a note is shown publicly (Mohammadi et al., 10 Oct 2025).

The contributor architecture is explicitly tiered. New Contributors begin by rating existing notes, and they can write notes only after their rating score reaches a threshold of 5. The system also maintains writing impact and rating impact scores, so note production and note evaluation are both part of contributor standing. Users may additionally request a note for a post, and if enough requests accumulate, high-impact contributors may be notified to write one. The platform’s data and code are described as open-source, and the literature has repeatedly relied on the public Community Notes repository for measurement and replication (Mohammadi et al., 10 Oct 2025).

This design places Community Notes between expert moderation and unrestricted user commentary. It is neither a conventional professional fact-checking program nor a simple voting widget. A plausible implication is that the system’s institutional identity depends on two linked claims: that note production can be decentralized, and that note visibility can still be constrained by an algorithmic notion of cross-perspective agreement.

2. Moderation workflow and scoring logic

Ordinary users can propose notes on potentially misleading posts and rate whether proposed notes are helpful. X then uses a bridging-based matrix factorization algorithm to score notes based on ratings from users with diverse viewpoints; only notes that clear a helpfulness threshold are attached and shown with the post. Early versions of the system were described as vulnerable to polarized rating and partisan behavior because the most-helpful-rated note could be surfaced regardless of who rated it. The bridging algorithm was introduced to require positive ratings from contributors with opposing viewpoints rather than from a single bloc (Slaughter et al., 18 Feb 2025, Mohammadi et al., 10 Oct 2025).

In the formal scoring model reproduced in later work, a rating from rater $u$ on note $n$ is modeled as

$$\hat{r}_{un}= \mu + i_{u} + i_{n} + f_{u} \cdot f_{n},$$

where the note intercept $i_n$ is used as a measure of helpfulness and the note factor as a measure of polarization. Community Notes statuses are dynamic: notes begin in Needs More Ratings, can become Currently Rated Helpful and be displayed publicly, and can later revert to Needs More Ratings or Currently Rated Not Helpful if subsequent ratings alter the inferred consensus (Chuai et al., 20 Jan 2026).

The workflow has also expanded beyond the original human-only pipeline. The “Request Community Note” function allows any user with a verified phone number to request a note on a post; once a post reaches five requests, it becomes visible to top writers, who may choose to write one. More recently, Collaborative Notes introduced an LLM-drafted initial note that is iteratively refined using human ratings and free-text suggestions, with the draft cycling until it stabilizes or locks after two weeks (Chuai et al., 12 Sep 2025, De et al., 29 Jun 2026).

3. Causal evidence on engagement and diffusion

Empirical work has not produced a single effect size because studies use different intervention definitions—platform roll-out, threshold-adjacent note display, matched DiD on repost time series, and post-level synthetic controls—and different outcome windows. The later causal literature nonetheless converges on a common pattern: once a note is visible or attached, subsequent engagement usually declines, but overall lifetime reductions are smaller because notes often arrive after early diffusion.

Study	Design	Main finding
(Chuai et al., 2023)	DiD and RDD on roll-out and visibility threshold	No robust significant reduction in retweets or likes; note visibility remained slow, with delay from tweet creation to note visibility falling from 2.85 days in the U.S. phase to 2.23 days globally
(Renault et al., 2024)	Staggered DiD near the helpfulness threshold	Visible notes reduced retweets by 49.1%, replies by 32.4%, and quotes by 34.6%; tweets with helpful notes were 80% more likely to be deleted; overall retweet virality fell by 16.34%
(Chuai et al., 2024)	Matched DiD on 237,677 repost cascades	Displaying notes reduced subsequent reposting by 62.0% and increased deletion odds by 103.4%; cumulative reposts fell by 15.3%; average time from post creation to note display was 61.4 hours
(Slaughter et al., 18 Feb 2025)	Synthetic controls on 40,074 posts with proposed notes	Forty-eight hours after attachment, reposts fell by 45.7%, likes by 43.5%, replies by 22.9%, and views by 14.0%; over full post lifespans, reposts fell by 11.4%, likes by 13.0%, replies by 7.3%, and views by 5.7%; cascades became less deep and less structurally viral, but not less broad conditional on size

The synthetic-control literature adds structural detail beyond raw engagement counts. In reducing reposts, Community Notes altered cascade geometry: fact-checked cascades were less deep and less structurally viral, while breadth no longer differed much once total repost growth was matched to synthetic controls. This contrasts with prior observations that false information diffuses farther than true information while remaining structurally similar conditional on reach, and suggests that moderation labels may change not only cascade size but the manner of diffusion (Slaughter et al., 18 Feb 2025).

Timing is the recurrent limiting variable. In one synthetic-control analysis, the first quartile of attachment speed produced a $-49.4\%$ growth reduction in reposts, while very late notes at 47+ hours were close to zero. Across the causal literature, this makes the system’s measured effectiveness highly sensitive to whether the estimand is “after note display” or “over the entire life of the post.”

4. Participation inequality, consensus, and temporal dynamics

Large-scale descriptive work depicts Community Notes as a highly skewed contributor ecology. A four-year resource dataset covering January 23, 2021 to January 23, 2025 reported 227,702 unique Contributors, 1,614,743 Notes, and Notes written on 1,016,673 distinct Posts. Participation was strongly concentrated: one Contributor authored 33,186 Notes, and the discussion notes that this account appeared automated and mainly targeted impersonation, NFT, and cryptocurrency scam accounts. Over the same interval, only 13.55% of Posts with at least one proposed Note ever received a helpful Note, 87.7% of all Notes remained in Needs More Ratings, only 8.3% ultimately became helpful and were displayed, and the average delay to helpful publication was 26 hours (Mohammadi et al., 10 Oct 2025).

Subsequent work sharpened the inequality and dissensus picture. In a corpus of approximately 1.8 million notes, the top 1% of contributors wrote 28.2% of all notes and the top 10% wrote 58.4%, with Gini coefficient $= 0.679$, Theil index $= 1.442$, and normalized entropy $= 0.884$. Consensus was rare in two senses: about 69% of posts with more than one note exhibited classification dissensus, and only 11.5% of notes reached agreement on publication while 88.5% remained in NMR. The same study reported that approximately 68% of posts were annotated as “Note Not Needed,” and that NNN-flagged posts were paradoxically more likely to yield published notes, with $OR = 3.124$ and $p < 10^{-300}$ (Razuvayevskaya et al., 14 Oct 2025).

Consensus is also unstable after publication. Among 43,782 displayed notes in a large observational dataset, 30.2% later lost helpful status and disappeared. Interrupted time series models showed that note display triggered a 39.3% immediate increase in rating volume and systematic shifts in rating leaning: contributors similar to note authors became more supportive, while dissimilar contributors became more negative. Counterfactual analyses attributed a substantial share of disappearance to this post-display polarization, especially from dissimilar raters (Chuai et al., 20 Jan 2026).

Experimental work on note authorship suggests that the crowd’s productive capacity is not exhausted by the current write-then-rate pipeline. In an online experiment on political posts, collaboratively authored two-person notes were rated as more helpful than individually written notes, but the collaborative benefit diminished when team members were explicitly shown one another’s political affiliations. This suggests that collective intelligence gains depend not only on diversity, but also on whether identity cues intensify partisan friction (Juncosa et al., 29 Jan 2026).

5. Fact-checking ecosystems and multilingual variation

A central debate concerns whether Community Notes can replace professional fact-checkers. The literature answers this negatively. In an English-language corpus of 664,000 notes, at least 5% of all Community Notes contained an external link to a professional fact-checker, rising to 7% among notes rated helpful, while only 1% of not-helpful notes contained a fact-checking source. The same study found that notes tied to broader misinformation narratives or conspiracy theories were about twice as likely to reference fact-checking sources as other claims, and argued that community moderation depends on the research infrastructure and investigative labor of professional fact-checking organizations rather than substituting for them (Borenstein et al., 19 Feb 2025).

Cross-linguistic work shows that this dependence is unevenly distributed. In 671,358 notes written between September 2024 and June 2025, notes appeared in 79 different languages, but the top five—English, Spanish, Japanese, Portuguese, and French—accounted for over 90% of all notes. English alone accounted for 413,950 notes. At the lower end, 34 languages had 10 or fewer notes, 15 languages had only one note, and among languages with fewer than 100 notes, 42 of 48 had no notes rated Helpful. Uptake was therefore highly uneven across linguistic communities (Stewart et al., 23 Dec 2025).

Professional fact-checking citation is also sparse but measurably associated with better note outcomes in multilingual data. Using IFCN signatories as the reference set, 21,530 notes, or 3.2% of all notes, cited an IFCN source. IFCN-citing notes were rated Helpful at 12.7%, compared with 8.1% for non-IFCN notes, and rated Not Helpful at 1.2%, compared with 3.8% for non-IFCN notes. Yet the overall rate remained low, and many communities relied on cross-language sourcing rather than same-language fact-checking infrastructures (Stewart et al., 23 Dec 2025).

This body of work supports a specific interpretation of Community Notes’ epistemic position. It is scalable only where a sufficient volunteer base, rating activity, and supporting source ecosystem coexist. Where those conditions weaken—whether because of low linguistic uptake or thin professional fact-checking coverage—the algorithm has little material to elevate.

6. Requests, AI augmentation, and automated note production

The request mechanism was introduced as a scaling device, but its effects are selective rather than comprehensive. In a study of 98,685 requested English-language posts and 5,888,351 requests, 53.6% of requested posts received a note, yet only 12.1% were estimated to be truly request-fostered—written by top writers after the fifth request threshold had been reached. Request-fostered notes were more helpful than writer-only notes and less polarized than other notes on requested posts, but contributors and requestors did not target the same material: contributors favored posts with higher GPT-estimated misleadingness and higher misinformation exposure, while requestors emphasized political content more heavily (Chuai et al., 12 Sep 2025).

Research on human–AI hybridization extends this scaling logic into note generation itself. A conceptual proposal termed Reinforcement Learning from Community Feedback argues for an open ecosystem in which both humans and LLMs write notes, but only humans rate them and determine which notes are shown. Empirical work on Collaborative Notes then studies one concrete implementation: 19,146 collaborative note versions on 10,600 unique posts, together with 211,850 instances of human feedback. Human feedback improved AI drafts, but collaborative notes still reached helpful status at lower rates than human-only or AI-only notes—1.8% versus 9.0% for human-written notes and 14.1% for AI-written notes—and attracted fewer raters. At the same time, collaborative notes were complementary rather than merely duplicative: 76.5% of posts with at least one collaborative note had no human-only or AI-only notes proposed (Li et al., 30 Jun 2025, De et al., 29 Jun 2026).

Fully automated note writing is no longer hypothetical. In the first eight months of the AI Note Writer API, 20 AI writers accounted for 14.2% of all submitted notes, and their daily share rose to 44.8% lately. AI writers were highly responsive, with median response time of 5.3 hours from post creation and 0.14 hours from API eligibility to note submission. They also expanded coverage: AI notes appeared on 16.8% of fact-checked posts, and 74.4% of those posts were not checked by humans. Their quality profile was mixed: AI-generated notes were less likely to be classified as helpful than notes written by human experts, though they outperformed notes written by laypeople (Gong et al., 15 May 2026).

A parallel methodological literature focuses on AI as note generator or note evaluator rather than direct platform actor. GitSearch reported 99% coverage on PolBench and a 69% win rate against human-authored helpful notes; CrowdNotes+ for health misinformation reported a median delay of 17.6 hours before the first note receives a helpfulness status and introduced HealthJudge with 81.03 Macro-F1 and 81.44 Macro-Accuracy; MultiCom framed note evaluation as persona-guided multi-agent simulation and reported 84.7% accuracy, 68.3% balanced accuracy, and 60.1% macro-F1 on note-status prediction (Singh et al., 9 Feb 2026, Wu et al., 13 Oct 2025, Wen et al., 3 Jun 2026).

7. Vulnerabilities, controversies, and design implications

The most persistent controversy is not whether Community Notes can reduce diffusion after a note appears, but whether it can do so early enough. Multiple studies identify delay as the dominant implementation bottleneck. One roll-out analysis noted prior evidence that roughly 95% of tweets stop getting relevant impressions after two days and that the median half-life of a tweet is about 80 minutes; later work reported average time from post creation to note display of 61.4 hours, and another large-scale efficiency study found that the first helpful note appeared after 65.7 hours on average and 15.3 hours at the median. This suggests that even strong post-display effects can translate into much smaller cumulative effects when early virality has already passed (Chuai et al., 2023, Chuai et al., 2024, Razuvayevskaya et al., 14 Oct 2025).

A second controversy concerns robustness to bias and manipulation. Simulation-based stress testing of the open-source algorithm found suppression of genuinely helpful notes consistently above 40% across conditions and showed that the system is highly sensitive to rater polarization and in-group preferences. With about 12% indiscriminate bad raters, or with 5–20% coordinated bad raters depending on the surrounding bias environment, targeted helpful notes could be strategically suppressed. Observational work on displayed notes complements this result by showing that post-display polarization can destabilize previously surfaced notes (Truong et al., 4 Nov 2025, Chuai et al., 20 Jan 2026).

A third debate concerns what Community Notes can and cannot substitute for. The literature does not support a simple replacement thesis—neither for professional fact-checkers nor for human raters. Professional fact-checking remains structurally important for high-stakes narrative-linked misinformation, while AI systems improve speed, coverage, and drafting capacity but currently underperform human experts on note helpfulness. The resulting picture is of a hybrid moderation system: effective in many settings, scalable relative to expert-only review, but constrained by timeliness, participation inequality, consensus instability, linguistic unevenness, and susceptibility to polarized rating behavior (Borenstein et al., 19 Feb 2025, Gong et al., 15 May 2026).

Overall, Community Notes is best understood as a consensus-based contextual moderation system whose empirical strengths and weaknesses are inseparable from its socio-technical design. The strongest evidence shows that attached or displayed notes can materially reduce reposts, likes, replies, views, and cascade depth, yet the broader literature equally shows that those gains are mediated by who participates, how fast notes appear, what evidence ecosystems contributors can draw on, and how resilient the scoring process is to organized disagreement (Slaughter et al., 18 Feb 2025).