Community Notes Explained: Crowd Moderation for a Feed

Community Notes is a crowd moderation tool: approved contributors write short notes that add context to a post, other contributors rate those notes, and a note is shown only when people who usually disagree both find it helpful. It does not delete posts or ban accounts. It adds a layer of context on top of normal moderation.

This guide explains the mechanism from the program's own open-source documentation and code, then shows where crowd notes fit in a moderation stack and how an operator of a smaller network could run a version. If you are planning a short-post network, a Twitter clone script gives you the feed, report queue and operator console to build that stack on. We describe the program as the repository documents it as of October 2026; thresholds and rules in it change, so check the current text before you quote a number.

What crowd notes are

The Community Notes repository describes a program that empowers people on X to collaboratively add helpful notes to posts that might be misleading. Four ideas from the introduction carry the design.

• Contributors drive it. Contributors are people on the platform who sign up to write and rate notes.
• Only notes rated helpful by people from diverse perspectives appear. The documentation says it does not work by majority rules.
• The company does not choose which notes show. The documentation says the platform does not rate or moderate notes unless they break its rules, and that only contributors' ratings determine which notes show.
• It is open. All notes, ratings and contributor data are published daily, and the ranking code is public.

The life of a note

1. A contributor sees a post that may mislead and writes a note.
2. Other contributors rate the note: helpful, somewhat helpful or not helpful, with reasons.
3. The ranking algorithm runs at intervals and assigns a status.
4. If the status is Helpful, the note appears on the post. Otherwise it stays hidden from general readers.
5. The post's author can ask for additional review. The documentation says that review is done by other contributors, not by staff.

Who contributes, and how they earn the right

A crowd program depends on who is in the crowd. The signing-up page lists eligibility: no recent notice of rule violations, an account at least six months old, and a verified phone number from a trusted carrier that is not tied to other Community Notes accounts. It explains the aim, which is to make accounts more likely to be real people and not adversarial actors. When more people apply than there are places, it admits applicants at random from country-specific waitlists, to avoid a crowd drawn from one ideology or interest.

New contributors can rate straight away but cannot write. The writing-ability page says they must first earn a Rating Impact of at least 5, which comes from rating notes in a way that matches how notes end up. Writing ability can be locked again if recent notes keep reaching Not Helpful, and the unlock threshold rises each time.

Contributions are anonymized under aliases but visible to the public. That choice supports the transparency goal: outsiders can check the data without learning who wrote what.

The bridging idea in plain language

Most rating systems add votes. Whoever has more supporters wins, which means a large, united group can push any note up or down. Community Notes asks a different question: do people who normally disagree both think this note is helpful?

The diversity of perspectives page explains how it tells people apart. It does not use demographics, location, follows or posts. It uses only past ratings: contributors who tend to rate the same notes the same way are treated as similar, and contributors who rate differently are treated as different. If people on both sides of that divide rate a note helpful, that is a good sign the note is useful to people across viewpoints.

A toy example

These numbers are invented. Say a network has two groups of contributors who usually disagree, Group 1 and Group 2, of equal size.

A simple vote would show Note A, which one side loves and the other side rejects. Bridging holds it back. This is the whole point: the system prefers a note that is mildly agreed by both sides over one that is enthusiastically backed by one.

What the code does with it

The ranking page describes matrix factorization on a note-by-rater matrix. Each rating is predicted from a global term, a rater term, a note term and the product of rater and note viewpoint factors. The model is built so that the viewpoint factors absorb as much of the pattern as they can, and the note's own term, which becomes its helpfulness score, rises only if raters with different viewpoint factors all rated it helpful. That is the math version of the table above.

Note statuses and timing

The ranking page names three statuses: Needs More Ratings, Helpful and Not Helpful. It says every note starts as Needs More Ratings and cannot move until it has at least five ratings. It sets a score of 0.40 and above for Helpful, and says that, at the time of writing, only notes that mark a post as potentially misleading and earn Helpful are eligible to appear on posts. Not Helpful uses a separate score cutoff and requires ratings from raters on both sides of the viewpoint scale. Treat these numbers as the documentation's, dated, and likely to change.

Two timing rules matter for any copy of the system:

• Statuses are computed periodically. The documentation says the delay lets the system collect independent ratings from people who have not yet been influenced by seeing a status on a note.
• Data is released with a lag. Contributor scoring counts as valid only ratings made before the rater could know the final status, which the documentation ties to the public release of rating data after 48 hours.

The point for an operator is that speed is a design choice. A system that shows a note the moment three friends rate it is fast and easy to game. A system that waits for diverse ratings is slower and harder to game.

Reputation: how contributors earn weight

The contributor scores page describes three scores. An author score compares the share of a contributor's notes that reached Helpful with the share that reached Not Helpful, with Not Helpful counted far more heavily. Another author score averages note scores. A rater score reflects how closely a contributor's early ratings matched the eventual outcome of notes where raters reached clear consensus. A minimum number of ratings is required before scores are computed, and contributors with scores that are too low are filtered out of the second round of scoring.

The effect is that a track record counts. Someone who rates carefully has more influence than someone who rates everything the same way. And deleting a bad note does not erase its effect on the author's record, which closes an obvious loophole.

Strengths and limits

Strengths

• Scales with volume. More posts need more contributors, not more employees.
• Adds context without removal. The post stays up and the reader sees both.
• Auditable. Public data and code let outsiders test it. The documentation also describes monitored quality measures, such as accuracy reviews by professional reviewers and surveys, which can trigger guardrail procedures.
• Harder to push through one-sided campaigns. Bridging is designed to resist this.

Limits

• Slow. A note needs ratings from enough diverse contributors. On a fast-moving post the context can arrive after the harm.
• Coverage gaps. Posts that few contributors see get no notes. Topics where the crowd cannot agree get none either, by design.
• Depends on active contributors. If contributors lose interest or leave, the system stalls.
• Narrow scope. The documentation says only notes on potentially misleading posts are eligible to display today. It does not cover spam, harassment, impersonation or illegal content.
• No legal role. A note does not satisfy a takedown notice, a court order or a store policy. Those need staff action.

Where it fits in a moderation stack

The response times are typical ranges we use for planning, not measurements. The right mix depends on your content. A network for professionals sharing research might rely more on crowd context. A network with younger members needs strong staff review and filters first. Our guides on content moderation models and moderating an online community at scale cover the staff and volunteer layers in more detail.

Could a smaller network run one?

Yes, with limits. A network with a few thousand members cannot reproduce the statistics of a very large one, because bridging needs enough contributors in different viewpoint groups to find agreement. But the principles carry over, and the code is public to read. Check the repository's license and notices before you reuse any of it.

Design steps for a light version

1. Decide the scope. Start with one job: context on posts that make factual claims. Leave harassment and spam to the report queue.
2. Choose contributor requirements. Copy the shape: an account age, a verified contact, no recent rule breaks. Pick numbers that fit your size and write them down.
3. Let people rate before they write. Require some rating activity before a member can write a note, as the documentation does. It teaches the standard and filters out the careless.
4. Define the rating inputs. Use a small scale (helpful, somewhat, not helpful) and a short set of reason tags, so you can tell why a note was rejected.
5. Set a status rule. Require a minimum number of ratings and agreement across at least two groups of raters who have disagreed before. With few members, define groups from past rating behavior, as the documentation does, not from profiles.
6. Delay the status. Compute statuses on a schedule instead of instantly, so early raters are not influenced by a visible result.
7. Keep a record. Store every note, rating and status change with a timestamp so you can explain a decision later.
8. Add an appeal route. Let the author of a post request a second look from other contributors, and let anyone report a note that breaks your rules.
9. Pilot privately. Run it for an invited group first, with staff watching how often they disagree with the outcomes.

Abuse prevention

• Sock puppets. Tie contributor status to a verified contact and an account age, and look for contributors who always rate together.
• Pile-ons. Require disagreement-spanning agreement before a note shows, and weight raters by track record.
• Gaming reputation. Count only ratings made before a status was visible, and keep deleted notes in the author's record.
• Automated writers. If you allow software to propose notes, treat it as an untrusted contributor with no special weight.

Pair it with auditable staff actions

Crowd context works best next to staff decisions that leave a trail. When a note is wrong, or a post needs removal, a moderator should be able to act in proportion: hide, restore, remove, suspend or ban, with a written reason. The platform we sell includes a report queue with five graded actions and an audit log that stores the administrator, target and reason for every privileged action. Those records answer the question that follows any dispute: who decided, and why. See the Twitter clone features for the full list.

The same evidence trail supports your Twitter clone business model if you sell paid tiers: members who pay expect fairness and speed, and records are how you show it. For the feed side of the stack, how the Twitter algorithm works explains how ranking and safety filters meet, and X vs Bluesky vs Threads compares crowd, staff and labeler models. If you are starting from nothing, read how to start a microblogging platform first.

Other kinds of networks have the same needs in different clothes. A TikTok clone app leans on automated screening for video, a ReelShort clone reviews licensed content before release, and a white-label OnlyFans clone needs identity checks and consent records. In each case crowd context would be a small layer, not the foundation.

How to tell whether it is working

The program's evaluation page describes three top-line quality measures: accuracy of notes that reach Helpful, how well they inform readers, and whether readers outside the contributor pool find them helpful. They are used for monitoring and do not change the status of any single note. A small network cannot run professional reviews or national surveys, but it can track cheaper versions of the same three questions.

The thresholds in the last column are planning prompts of ours, not figures from the program. Review them monthly in the first quarter, and stop the pilot if the sample shows notes that mislead. A white-label Twitter clone with an audit log makes the last two measures easy, since every hide and correction is already stored with its reason.

Checklist before you pilot crowd notes

• Staff moderation, a report queue and an audit log work already.
• The job of notes is written in one sentence.
• Contributor requirements are written and tested on a small group.
• Ratings use a short scale and reason tags.
• Status needs a minimum number of ratings and cross-group agreement.
• Status is delayed and computed on a schedule.
• Every note, rating and status change is stored.
• Authors can request a second review.
• Someone measures how often staff overrule the crowd.
• Your terms say notes are not legal advice or a takedown substitute.

Short glossary

• Note. A short piece of context a contributor attaches to a post.
• Bridging. Requiring agreement from people who usually disagree.
• Needs More Ratings. The starting status of every note.
• Rating Impact. A measure of how well a contributor's ratings match final outcomes.
• Alias. A public pseudonym a contributor uses for notes and ratings.
• Guardrail. A quality measure that can trigger a response when notes look unreliable.

What to decide next

Decide first whether context is the job you need done. If most of your problems are spam, harassment or illegal content, build the staff queue before anything else. If misleading claims are your main concern and you have an active, trusted membership, a small pilot of crowd notes is a reasonable addition. Compare the staff tools on the pricing page, see how it works, and bring your plan to us when you are ready to scope it.