X Community Notes (crowd annotation)

A post goes up on X. Somebody thinks it is misleading. In every other deployment in this domain what happens next is that a classifier scores the post, a reviewer looks at a queue, and something is removed, restricted or downranked. Here none of that happens. A volunteer writes a note — a paragraph of context with, in 79.6 percent of English notes, at least one cited source URL — and proposes it. Other volunteers rate that note helpful, somewhat helpful or not helpful, with structured tags. Every hour, a matrix-factorization model is re-trained from scratch over the whole rating matrix. If the note's fitted intercept reaches 0.40 with a latent factor magnitude under 0.50, the note appears beneath the post. The post itself is untouched.

The operator says so in terms. Its published FAQ states that notes rated helpful by enough contributors from different points of view "will appear directly on posts" and that "beyond that, notes do not affect display of posts or enforcement of X's Rules." Its introduction page says "X doesn't write, rate or moderate notes," except where a note itself breaks the platform's rules, and that the system "doesn't work by majority rules": a note requires agreement between contributors who have sometimes disagreed in their past ratings. The company's stated role is to run the scorer and publish the result.

That bridging requirement is a published parameter choice rather than a slogan, and it is worth reading as arithmetic. The scorer fits a global intercept, a per-rater intercept and factor, and a per-note intercept and factor. Intercept terms are regularized at 0.15 and factor terms at 0.03 — five times more strongly — which the operator's own ranking documentation says is what "requires that notes are rated by raters with diverse factors before a note gets a label." A note needs at least five ratings even to leave the "Needs More Ratings" state; a tag-outlier filter raises the display bar from 0.40 to 0.50; and status changes are deliberately delayed before they are shown, so that a rater's judgement forms without knowing where a note already stands.

The corpus is public. Five files — notes, ratings, note status history, contributor enrolment, note requests — are released daily on a best-effort basis, cumulative, containing items created up to 48 hours before release, under participant identifiers that are stable across handle changes. The scoring code sits in a public repository under Apache-2.0. This is why the deployment has an independent causal evidence base at all: three separate research teams estimated its effects from the operator's own data without asking the operator's permission, and none of that is possible for any classifier-based enforcement system in this atlas.

Start with what the mechanism gets right, because it is measured and it is substantial. A synthetic-control study of 40,078 posts for which notes were proposed between 16 March and 23 June 2023 — of which 6,757, or 16.9 percent, received a helpful note — estimated post-attachment growth reductions of 46.1 percent in reposts, 44.1 percent in likes, 21.9 percent in replies and 13.5 percent in views, with the noted content's repost cascades becoming less deep and less structurally viral than matched counterfactuals. An independent difference-in-differences study of 237,180 cascades reposted more than 431 million times reached the same shape by a different identification strategy: 61.2 percent less subsequent spread, and 94.3 percent higher odds that the author deletes the post. And the notes themselves are accurate. In a randomly sampled set of notes on popular COVID-19 vaccine posts, evaluated with an infectious-disease physician and a virologist, 97.5 percent were entirely accurate, 2 percent partially accurate and 0.5 percent inaccurate; 93 percent cited sources of high or moderate credibility.

Now read the same two studies' other numbers. The synthetic-control study also measured its effect over the whole life of the post, and there it collapses to 11.6 percent fewer reposts, 13.3 percent fewer likes, 6.9 percent fewer replies and 5.5 percent fewer views. The difference-in-differences study reports a system-wide reduction in total engagement with misleading posts of 14.9 percent, and states the reason: notes "often appear too late to intervene in the early (and most viral) stage of the diffusion." The gap between each pair of numbers is the engagement that had already happened before the note appeared.

The dose-response on speed is measured and monotone. Notes attached within 1 to 12 hours cut lifetime reposts by 24.9 percent; at 12 to 23 hours by 12.3 percent; at 23 to 47 hours by 4.3 percent; beyond 47 hours by 0.1 percent, statistically indistinguishable from nothing. In that slowest quartile view growth and reply growth actually rose, by 13.6 and 27.0 percent, consistent with a late note dragging attention back to a stale post. Against that clock: roughly 50 percent of a post's reposts happen in its first five hours and 80 percent within sixteen, and the median half-life of a post's impressions is about 79.5 minutes. There is no single figure for how long a note takes. Published central estimates run from a 14.3-hour median in one 2021-2023 sample, through roughly 23 hours implied by the 2023 speed quartiles, 24.29 and 26 hours in two later reports, and means of 2.23 to 2.85 days by rollout period, to a 65.7-hour mean over 1.8 million notes — different clocks, different samples, different periods. The fastest display observed anywhere in one large dataset was 80.2 minutes. The earliest rigorous evaluation, a difference-in-differences and regression-discontinuity analysis of the US and worldwide roll-outs, found no evidence that introducing the system reduced aggregate engagement with misleading posts at all, and attributed that null to display latency exceeding the diffusion half-life. That is not a contradiction of the later post-level work. They measure different things, and the discrepancy is the finding.

Then there is coverage, which is the larger half. An independent parse of the whole public corpus from 23 January 2021 to 23 January 2025 counts 227,702 contributors writing 1,614,743 notes on 1,016,673 distinct posts — and finds that 87.7 percent of those notes stayed in "Needs More Ratings" and 8.3 percent reached the helpful state. Only 13.55 percent of posts with at least one proposed note ever received a helpful one. Other samples agree in magnitude: 11.3 percent, 11.5 percent, "about 10 percent and declining", and a newspaper count of roughly 79,000 of more than 900,000 notes written in 2024 shown publicly, under 9 percent.

And the gate fails hardest exactly where the stakes are highest. The Center for Countering Digital Hate reported on 30 October 2024 that 209 of 283 sampled misleading US-election posts — 74 percent — had accurate notes written about them that were never shown to all users, with the topics on which cross-perspective agreement most often failed being claims that the 2020 election was stolen, claims about imported illegal voters, and claims about voting-system safety. That is advocacy research on a purposive, non-random sample, and it is reported here as such; the publisher's site serves a captcha interstitial to automated requests, so the figures are carried through trade coverage that quotes them verbatim. It is corroborated in shape by the peer-reviewed work: the measured spread-reduction effect is significantly weaker for political content and for posts from influential accounts, and an archival analysis of more than 1.8 million notes finds 69 percent of noted posts receiving classifications from contributors that conflict with one another. A mechanism that publishes only what raters who normally disagree can agree on is, by construction, weakest on the claims people disagree about.

The crowd behind it is unequal and largely monolingual. The top 10 percent of contributors wrote 58 percent of all notes, a Gini coefficient of 0.68; one apparently automated account wrote 33,186 of them; and although notes appear in 103 languages, only about 16 percent of note authors ever write in more than one, so the system is a set of partially separated single-language crowds rather than one crowd. Admission is by three published criteria — six months of account age, a verified phone from a trusted carrier not associated with another contributor account, and no recent notice of rule violations — with random selection from country-specific waitlists where applicants exceed slots. The publication bottleneck feeds back on that population: a regression-discontinuity analysis finds that having a note published raises the retention of first-time contributors, so a low and declining publication rate is a threat to the system's own labour supply and not only to its output.

Since 1 July 2025 the writing side has admitted machines. Third-party automated note-writing agents may propose notes through a published interface after passing a test-mode screen — at least 95 percent high scores on URL validity and at least 30 percent on whether the note addresses a claim rather than an opinion, over the most recent 50 test notes — with daily writing limits starting at 10 and scaling between 2 and 500 with measured helpfulness. Ratings stay human: the documentation states that "ratings come from regular contributors (i.e. humans) whose input ultimately determines which notes show." A think-tank evaluation of the pilot's first seven months, computed from the operator's public downloads, counts 27 accounts enrolled and 24 active, writing 31,464 notes — 7.4 percent of note volume but 13.9 percent of displayed notes, reaching the helpful state at 18.0 percent against 8.9 percent for human writers and consuming roughly 304 ratings per displayed note against 908. Its median time from note creation to a first non-unrated status was 6.0 hours against 6.3 for humans; that clock starts at note creation and is not comparable to the post-to-display latencies above. The evaluation measured no engagement outcome at all and says so.

Three more facts belong in the record. First, one monetary consequence sits outside the display channel and in tension with the operator's own documentation: on 29 to 30 October 2023 the platform's owner announced that posts carrying a note become ineligible for creator ad-revenue sharing, framing it as maximizing "the incentive for accuracy over sensationalism" and asserting that attempts to weaponize notes to demonetize people "will be immediately obvious" because the code and data are open. Claims of large algorithmic reach penalties for noted posts appear only on low-tier marketing pages, are supported by no credible source, and are not repeated here. Second, the visible layer is operationally fragile: around 22 to 26 May 2025 notes stopped appearing in users' feeds for several days after a data-centre outage, with the operator's engineering account acknowledging continuing issues and the Community Notes account saying on 26 May it was "working to get notes appearing normally". A very large platform's correction layer went dark as an operational side effect, with no policy decision and no announcement. Third, an agent-based evaluation of the published scorer reports that under polarization and in-group rating preference it suppresses a substantial fraction of genuinely helpful notes, and that a coordinated minority of 5 to 20 percent of raters could strategically suppress targeted helpful notes. That is a simulation of the published algorithm on synthetic data, not a measurement that manipulation occurred; no source here documents a successful real-world suppression campaign.

The mechanism has been copied, and that is now part of what it is. On 7 January 2025 Meta announced it would end third-party fact-checking in the United States in favour of a community-notes-style model; on 18 March 2025 it began testing on Facebook, Instagram and Threads with about 200,000 signed-up contributors, stating it would "use X's open source algorithm as the basis of our rating system" and that, unlike the fact-check labels it replaced, "notes will provide extra context, but they won't impact who can see the content or how widely it can be shared." On 26 March 2026 that company's Oversight Board issued a policy advisory opinion on expanding the model beyond the United States, concluding that "delays in note publication, the limited number of published notes and its dependence on the broader information environment's reliability raise serious doubts about the extent to which community notes can meaningfully address misinformation linked to harm," and recommending that countries with repressive human-rights records, imminent major elections, histories of coordinated disinformation networks, active crises, unsupported language complexity or persistent internet-access obstacles be omitted or delayed. That opinion is non-binding, it is addressed to Meta rather than to X, its evidence is X's own published performance figures, and it is context here rather than a ruling about this deployment. No adoption is described as complete or global, because as of that opinion the expansion beyond the United States was a plan under review.

Finally, the regulatory position, stated exactly. The European Commission's first-ever formal proceedings under the Digital Services Act, opened 18 December 2023, expressly put this mechanism under investigation: among the grounds is "the effectiveness of measures taken to combat information manipulation on the platform, notably the effectiveness of X's so-called Community Notes system in the EU and the effectiveness of related policies mitigating risks to civic discourse and electoral processes," against Articles 34(1), 34(2) and 35(1). That limb has produced no preliminary finding and no decision. The Commission's July 2024 preliminary findings covered the blue checkmark design, the advertisement repository and researcher data access; its 120 million euro non-compliance decision of 5 December 2025 rests on the same three grounds. Neither touches Community Notes. A Commission spokesperson declined to comment on the mechanism in May 2025 because of the ongoing proceedings, while confirming the investigation into its effectiveness. No regulator anywhere has published a finding on this system, and no litigation of any kind concerns it.