Source-linked AI summary
Community-Driven API and AI Writer Design for Openly Scaling Community Notes
Brad Miller, Jay Baxter, Jiansong Chao, Keith Coleman, Sophie Hilgard, Daniel Ortiz
TL;DR
Community Notes relies on contributor ratings to decide which notes show broadly, while proposing notes requires effort and can create latency. The paper presents an open AI Note Writer API and a community-guided writer; AI note writing adds coverage and improves speed and rating efficiency, with human contributors retaining authority over display.
Problem
Proposing notes requires contributor effort, which affects note supply and can create latency when addressing misleading information.
Source: §1 · p. 1Method
The paper presents a public AI Note Writer API and a Community Writer guided by community input to prioritize, draft, evaluate, and delete proposed notes.
Source: §1 · p. 1Results
Takeaways & Limitations
Human contributors remain the core and ultimate authority, governing display of both AI- and human-proposed notes.
Source: §8 · p. 8Comparison between writing techniques is limited because the study does not control for AI writers' operational costs or post-feed sizes and aggregates other AI writers.
Source: §7 · p. 8
Abstract
from arXiv · showhide
Community Notes is a crowd-sourced approach for adding context to posts on X. Contributors propose and rate notes, forming the inputs to an open-source, open-data algorithm that determines which notes show broadly to users. Since September 2025, Community Notes' AI Note Writer API has provided an open, public interface for using AI to propose notes, while adhering to the founding principle that users, not the platform or an AI, control which notes show on X. Explicit note requests and user posts on X determine the AI API post feeds, ensuring that AI note writing responds to demand from X users. We present the design, operation and impact of the AI API, including analysis of the interaction between AI and human generated notes across topics. Unless otherwise stated, measurements and system description reflect June 2-29, 2026. The Community Writer is the largest AI API client and contributes the bulk of AI API output, generating 52% of notes selected as Helpful and shown broadly on X. The writer is guided by community input during both training and operation to prioritize, draft, evaluate and delete proposed notes. Beyond scale, the writer also offers speed, submitting the first proposed, non-deleted note on 60% of posts when compared to other writers. AI note writing is additive on top of human note writers, extending coverage of Community Notes on X. Among posts that have Helpful notes, 42% have only AI notes, indicating human raters did not feel motivated to propose an alternative. In contrast, 30% have only human notes, reflecting contribution beyond the scope of AI writing. The Community Writer is open-source software released under the Apache 2.0 license.
1 Introduction
Community Notes uses contributor ratings and an open algorithm to surface notes that reach consensus among people who normally disagree. The paper introduces a public AI Note Writer API and the Community Writer, describing how they add note supply while leaving display decisions to contributors.
Background
- Community Notes relies on contributors to propose and rate context notes; an open-source, open-data algorithm identifies consensus among users who normally disagree, and CRH notes show broadly on X.Prior work found Community Notes accurate and effective at reducing sharing of misleading posts [3] [17].Source: §1 · p. 1
AI Note Writer API
- Since September 2025, the public AI Note Writer API has let clients consume candidate-post feeds and propose notes, while contributor ratings still determine which notes show broadly.Feeds reflect X user engagement, including explicit note requests and user posts; AI writers have helped increase daily CRH note volume by 86.0%.Source: §1 · p. 1
Community Writer
- The Community Writer is the largest AI API client, generating 41% of proposed notes, and combines specialized and commercial models to draft notes and filter candidates.It uses community input in its design, and its software is open-source under the Apache 2.0 license.Source: §1 · p. 1
Contributions
- AI API clients generated 52% of CRH notes while consuming 22% of ratings, and submitted the first non-deleted note on 60% of posts with a proposed note from another writer.The API clients' notes accounted for 55% of CRH note views.Source: §1 · p. 1
2 Background and Related Work
Prior work examines Community Notes accuracy and effects, related professional fact-checking, and AI fact-checks mostly outside production settings. This paper instead studies an open API and its largest client in platform operation.
Community Notes research
- Prior Community Notes studies report accuracy and reduced spread of misleading posts, while research also examines note content and community versus professional fact-checks.Reported findings include a 97% accuracy rate, a 46% repost drop [17], and a 62% average reduction in spread [3].Source: §2 · p. 2
AI fact-checking research
AI note writing
- The paper presents an open API for adding context across a platform and the design of its largest client, extending prior work focused on individual fact-checks.Its evaluation also measures latency, quality, and quantity across topics, including comparisons with top human contributors and other API clients.Source: §2 · p. 2
3 Community Notes AI API Overview
The AI Note Writer API lets clients propose, evaluate, submit, delete, and track notes through feeds shaped by user demand. Earn-in checks and outcome-based quotas govern publishing, while feed access and priority are tied to post signals and client performance.
API design
- The API exposes endpoints to request candidate posts, evaluate notes, submit or delete them, and retrieve real-time note status and ratings.Earn-in and quotas support contributors and clients, while candidate feeds are structured around X user note requests.Source: §3 · p. 2
Earn-in
- New clients must meet URL validity, abuse-classification, and ClaimOpinion quality checks on their 50 most recently submitted notes to earn publishing ability.The ClaimOpinion model predicts whether a note is likely to be Helpful by assessing whether it addresses a post's claims without imparting opinion.Source: §3 · p. 2
Quota system
- After earn-in, quotas rise with hit rate: 10% allows 50 notes per day, while 20% allows the maximum 500.Hit rate is calculated from CRH, CRNH, and NMR outcomes; the 14-day calculation excludes certain notes with fewer than 10 ratings.Source: §3 · p. 2
Post feeds
- Four feed sizes—Small, Large, XL, and XXL—use explicit requests, implicit engagement, and requester helpfulness; posts begin in XXL and move smaller as signals accumulate, while Small has the highest CRH-note concentration and latency.Source: §3 · p. 3
Feed access
- Clients qualify for feeds beyond Small with at least 100 notes written, a 5% hit rate, and at most a 10% CRNH rate; XXL access also requires net CRH of at least 100 over 90 days.Qualification requirements tie access to client outcomes.Source: §3 · p. 3, Table 1 · p. 3
3 Community Notes AI API Overview
- Of 65 client accounts that began earn-in, 32 earned writing ability and submitted a note for ratings, and 19 of those remained active.Among accounts that did not publish, 13 did not complete the 50-note earn-in, 5 failed ClaimOpinion checks, and 15 completed earn-in but never published.Source: §3 · p. 3
4 Community Writer Overview
The Community Writer uses community-derived signals and models to prioritize, draft, assess, revise, and delete proposed notes. Its design aims to respond to contributor and user demand while controlling latency and use of contributor ratings.
4 Community Writer Overview
- The Community Writer combines feature-based classifiers with fine-tuned and commercial models to prioritize, draft, evaluate, and delete notes using community input during training and inference.It operates separately from SpaceXAI infrastructure, uses only the public AI API, and faces the same earn-in, quota, and scoring rules as other writers.Source: §4 · p. 3
4.1 Design Goals
- The design targets responsiveness to contributor engagement and user demand, scalable low-latency writing, and efficient use of contributor ratings.Among NMR notes, 54% have fewer than 10 ratings, the effective minimum required for CRH status.Source: §4.1 · p. 3
4.2 Design Overview
- The writer queues fetched posts by feed size and Notable Post Model score, prioritizing smaller feeds and higher predicted CRH potential.Writers generate candidate notes using Grok 4.20 and CN Grok; models and rejectors assess candidates before submission, with a writer able to decline when context or sourcing is insufficient.Source: §4.2 · pp. 3–4
- Retry and revision queues reprocess selected posts after delays, while a real-time Deletion Model removes proposed notes unlikely to achieve CRH status.Retry follows an initial attempt within 45 minutes; revisions repeat after one and three hours for notes not yet rated CRH.Source: §4.2 · p. 4
5 Pipeline Component Details
The pipeline combines post-prioritization, note writing and rejection, and rating-based deletion to manage coverage and contributor-rating costs. Its components use community-note outcomes and post signals to prioritize posts, assess drafts, and remove unlikely-to-be-helpful notes.
5.1 Notable Post Model
- The NPM predicts whether a post will receive a CRH note using engagement, author-history, and post-context features.Training posts come from API feeds and are labeled by whether they received a CRH note, regardless of writing source.Source: §5.1 · p. 4
- The NPM records features as posts enter feeds and normalizes engagement by post age and impressions before a single-hidden-layer perceptron models interactions.This feature capture is intended to preserve temporally accurate training data.Source: §5.1 · p. 4
- The NPM is most accurate on larger feeds, where fewer posts are likely to receive CRH notes; Figure 3 evaluates predictions using notes from other writers, excluding CW notes.Framing recall measures posts without CRH notes that can be dropped, while false positives are pruned posts that ultimately receive a CRH note.Source: §5.1 · pp. 4–5, Figure 3 · p. 5
5.2 Writers and Rejectors
- CN Grok uses a brief prompt informed by Community Notes post-training, while Grok 4.20 receives detailed guidance to assess misleading claims and refine drafts.CN Grok's prompt includes user-suggested sources and an option to decline when a post is not misleading.Source: §5.2 · p. 5
- The rejectors screen drafts for predicted helpfulness, repeated mistakes, media-source mismatches, and whether revisions add claims, sourcing, or clarity.The Helpfulness Rejector becomes more permissive as user demand increases; Screenshot Rejector checks source screenshots against media in the post.Source: §5.2 · p. 5
5.3 Deletion Model
- Across feeds, the NPM, writers, and rejectors together yield similar CRH rates per post with one or more submitted notes.The NPM filters more posts on larger feeds while maintaining coverage of posts with human CRH notes.Source: Table 2 · p. 6
- The Deletion Model predicts note status from aggregated rating counts, using discretized features and polynomial crosses trained on snapshots from the first 30 ratings.The API aggregates ratings and tags into positive, neutral, and negative bins corresponding to rater-factor values; training extracts features at up to 10 intervals.Source: §5.3 · p. 6
- Deletion-model accuracy improves as ratings accumulate, and the CW polls every 60 seconds, applying the model from 3 ratings when confidence is high.The standard threshold is applied with 5 or more ratings.Source: §5.3 · p. 6, Figure 4 · p. 6
6 Evaluation
Across June 2–29, 2026, the Community Writer produced a disproportionate share of helpful notes and views while using fewer contributor ratings. Its feed pipeline expands coverage while balancing quality, speed, and human-AI complementarity.
6 Evaluation
- The XXL feed is ~57x larger than Small, raising H-CRH coverage from 24.3% to 73.4%, while median post age falls from 604 to 185 minutes.The XXL feed covers posts generating 8.1% of X post views daily.Source: §6 · p. 6
- The CW generated 41% of proposed notes, 52% of CRH notes, and 55% of CRH note views while consuming 22% of contributor ratings.Figure 5 shows its above-average CRH rate and impressions per CRH note alongside below-average rating use.Source: §6 · pp. 6–7, Figure 5 · p. 7
- NPM filtering passes 17.1% of XXL-feed posts and reduces H-CRH coverage from 73.4% to 55.8%; rejectors yield CRH rates of 7.9%–10.5% across feed sizes.Draft generation becomes less frequent as larger feeds contain a lower concentration of misleading posts.Source: §6 · p. 7
- Lower misleading-post density in larger feeds requires more precise prediction, so the CW raises evaluator and rejector thresholds and segments XL/XXL rejection by NPM confidence.NPM and rejection mechanisms accommodate class imbalance in larger feeds without exhausting note-writing quota.Source: §6 · pp. 6–7, Figure 6 · p. 7
- Conditioned on stable notes, the CW has the highest CRH rate despite more deletions, many below rating thresholds; its per-note quality compares well with other AI writers and humans.Stable notes include CRH, CRNH, deleted notes, and notes meeting rating minimums.Source: Figure 7 · p. 7, §6 · p. 7
- Among non-deleted notes on posts with another writer’s proposal, the CW submits first overall on 60% and first among AI writers on 95%.The paper links this lead to faster CRH status and increased views of CRH notes.Source: Figure 8 · p. 8, §6 · p. 8
- CW ratings lead to a CRH note 29% of the time, compared with 17% for notes by top writers.Outcomes per rating reflect the likelihood that contributor effort yields a CRH note.Source: §6 · p. 8
- Among posts with CRH notes, 41% have no AI-written CRH note and 30% have no AI proposal, while most AI CRH notes occur without a human proposal.The distribution spans 10 topics, assigned by Grok 4.Source: §6 · p. 8
7 Discussion and Limitations
AI note production adds to human writing while remaining community-guided and subject to human ratings. Rating-focused models aim to conserve volunteer attention, although their effects on retention are only potential; comparisons among writing techniques remain limited.
7 Discussion and Limitations
- Community input guides the CW’s feeds, training, and operation, while all notes remain subject to human raters; human notes continue to matter for rapidly evolving or nuanced topics.Human ratings demonstrate the utility of AI-generated notes and the ongoing value of human note writing.Source: §7 · p. 8
- The Notable Post and Deletion Models focus ratings on notes most likely to become CRH, saving volunteer effort; productive ratings may improve contributor retention and engagement.The paper describes rating a broadly displayed CRH note as potentially satisfying to contributors.Source: §7 · p. 8
- Comparisons between writing techniques are limited because operational costs and feed sizes are uncontrolled, and other AI writers are analyzed in aggregate.Writer segmentation is difficult because accounts may change, use multiple accounts, or combine writing behaviors.Source: §7 · p. 8
8 Conclusion
The AI API and Community Writer expand Community Notes coverage while improving writing speed, rating efficiency, and the yield of Helpful notes. Human contributors remain the ultimate authority over which notes show broadly.
8 Conclusion
- The AI API and Community Writer offer an open approach to increasing Community Notes coverage on X.The API and writer improve speed, rating efficiency, and the yield of notes rated Currently Rated Helpful and shown broadly.Source: §8 · p. 8
- AI note writing complements notes from human contributors, improving speed, rating efficiency, and the yield of Currently Rated Helpful notes shown broadly.Source: §8 · p. 8
- Human contributors remain the core and ultimate authority governing display of both AI- and human-proposed notes.Source: §8 · p. 8
A API Quota Algorithm
The progressive writing quota tracks recent note outcomes and writing volume alongside a daily writing limit.
A API Quota Algorithm
- The quota records the number of CRNH notes among the last 5 and 10 CRH/CRNH notes.Source: A API Quota Algorithm · p. 9
- It tracks the CRH-minus-CRNH rate over the last 20 and 100 notes and over 14 days among notes with at least 10 ratings and CRH or CRNH status.Source: A API Quota Algorithm · p. 9
- It also tracks average daily notes written over the last 30 days, total notes ever submitted, and the daily writing limit.Source: A API Quota Algorithm · p. 9
B Ratings per Status Outcome
Many notes needing more ratings do not reach the rating threshold associated with CRH status, while contributor ratings remain the sole mechanism for broad display.
B Ratings per Status Outcome
- 54% of NMR notes receive fewer than 10 ratings, effectively preventing them from reaching CRH status.Figure 10 covers non-deleted CRH, CRNH, and NMR notes and shows rating counts at first CRH or CRNH status attainment.Source: Figure 10 · p. 9, B Ratings per Status Outcome · p. 10
- Notes generally require at least 10 ratings for CRH status and at least 5 for CRNH status; contributor ratings alone allow notes to show broadly.Source: B Ratings per Status Outcome · p. 10
- Low rating counts reflect contributors’ freedom to propose notes on low-visibility posts and decline to rate notes they do not find helpful.Source: B Ratings per Status Outcome · p. 10