Community Notes
Community Notes is Twitter/X’s crowd-sourced fact-checking system. Contributors rate misleading or false posts and write contextual notes; notes that meet a threshold of cross-partisan agreement appear publicly beneath the relevant post. The system was built by Keith Coleman (Product Lead) and Jay Baxter (Founding ML Engineer) and has operated across multiple languages at the scale of millions of rated notes.
The design problem Community Notes addresses is not simply “how do we identify false content” but “how do we build a fact-checking system that cannot be captured by any political faction.” The bridging algorithm is the answer to the second problem.
The bridging algorithm
Most content moderation systems aggregate approval by volume: the note with the most upvotes wins. This fails immediately in a polarised environment because well-organised ideological groups can outvote disorganised ones. A note favourable to one party’s worldview accumulates approvals from that party’s supporters regardless of its accuracy.
The bridging algorithm rejects volume as the primary signal. Instead, it weights notes by the diversity of their approval coalition. A note approved predominantly by users who typically agree with each other — who rate the same content the same way, who cluster together in the rating data — earns little weight regardless of total approval count. A note approved by users who typically disagree with each other earns high weight.
In practice: if a note about a political claim draws approval from users whose other rating behaviour places them on opposite sides of a partisan divide, that convergence is strong evidence the note is accurate rather than merely congenial. The algorithm surfaces the signal of cross-partisan agreement; partisan consensus counts for much less.
This is not a neutral-by-default system in the sense of refusing to make calls. It makes calls aggressively — it simply makes them on the basis of disagreement convergence rather than majority preference.
Transparency and open-source
The algorithm is published in full. Rating data — anonymised contributor IDs and note scores — is downloadable by anyone. External researchers can reproduce the scoring independently and verify that the system operates as described.
This transparency is not incidental. It is a structural safeguard: any systematic manipulation of Community Notes output — surfacing notes that favour one political direction, suppressing notes that disfavour it — would be detectable from the public data. The openness creates accountability that internal policy cannot provide, because internal policy can be changed without external visibility.
The open-source decision also enables external researchers to propose improvements to the algorithm and to audit specific controversial decisions. The system’s credibility depends partly on this external scrutiny being possible and on the company not being the sole arbiter of whether its own system works.
Survivability and institutional design
Community Notes survived the transition from Twitter to X under Elon Musk — a leadership change that reversed many of the platform’s prior content moderation policies. Coleman and Baxter attribute this partly to the algorithm’s design: a system that cannot be tuned to favour any political direction without the manipulation being visible in public data is harder to co-opt than a discretionary moderation policy.
The survivability test is meaningful because it shows the system was not merely tolerated under one regime but was resistant to replacement under a regime with different political sympathies. The open-source structure that makes external auditing possible also makes it costly to quietly change the system’s behaviour — any change to the algorithm is immediately visible to the research community that monitors the public data.
This is institutional design in the strict sense: embedding constraints into a system’s architecture that outlast the intentions of any particular leadership.