How X built Community Notes: crowdsourced fact-checking that actually works

Executive overview

Most misinformation tools either fail on scale, fail on speed, or aren't trusted because a tech company controls them. Community Notes sidesteps all three problems by letting the public write and rate notes — with the algorithm only surfacing notes that find agreement across people who normally disagree.

The result: 95,000 notes seen 30 billion times in 2024, a 50–60% drop in resharing after a note appears, and adoption by Meta as their primary fact-checking mechanism.

The core insight: surprising agreement between polarized people is a reliable signal of accuracy — and manipulation resistance comes for free.

How the algorithm works

• Someone proposes a note on any post; it enters a rating pool
• Other contributors rate notes; the system tracks which raters typically disagree
• A bridging-based agreement score (matrix factorization) measures how much people who normally disagree find the note helpful
• A note must clear a 0.43 threshold on this scale — calibrated to require a sizable majority on both sides of any polarized divide
• A secondary filter removes notes rated helpful but also tagged as factually incorrect
• Only ~8% of proposed notes ever go live; quality is the non-negotiable constraint
• Contributors who consistently back bad notes lose writing privileges until they earn them back

Why the algorithm beats alternatives

• Majority-vote approaches amplify whichever group has more users
• PageRank variants resist vote rings but can't correct for ideological bias in the rating graph
• An internal bake-off (Kaggle-style) among ML engineers surfaced the bridging approach after real pilot data showed bias was the primary attack vector, not bot manipulation
• The algorithm and all rating data are fully open-source; anyone can replicate the output on a single machine (~500 GB RAM, ~one day)
• Vitalik Buterin and other external researchers have independently verified the algorithm does what it claims

Scale and impact

• ~950,000 contributors globally; hundreds of notes per day vs. ~10 traditional fact-checks per day
• A single note on an image auto-matches to every post sharing that image — one note can reach thousands of posts
• External A/B tests: 30–40% drop in likes and reposts while a note is showing; 50–60% total repost drop in difference-in-differences analysis
• Authors become 80% more likely to delete their post after receiving a note
• Median time from post to live note: five hours; traditional fact-checking typically takes two to four days
• During the first three days of the Israel–Hamas conflict in October 2023, 500 notes appeared on out-of-context imagery and fabricated footage

Founding principles

• Voice of the people: no company button can change a note's status; if a bad note appears, it signals a system flaw to fix, not a note to suppress
• Open to all: any account with a verified phone number can apply; contributors are randomly selected, not curated
• Full transparency: code on GitHub, daily data dumps downloadable as TSVs; the architecture was deliberately designed around this requirement from day one
• Anonymity/pseudonymity: pilot data showed contributors were more willing to cross partisan lines and write notes on controversial topics when not posting under their real name — counter to the team's original assumption

The thermal team model

• One clear driver with founder-level ownership of the project
• One senior decision-maker as the single escalation path (Kayvon early on; Elon after acquisition)
• 100% focus: everyone on the team works on nothing else; a question asked in the morning can have results by afternoon
• Self-selected membership: no one was assigned to the team; every hire was interviewed by the team and opted in
• Lightweight process: goals set milestone-by-milestone, not quarterly; coordination via a single long-running Google Doc; no JIRA, no Asana for internal work
• The team started at five people — one each in backend, frontend, ML, design, and research

Lessons from surviving four leadership changes

• The product's bridging-based design means leaders across the political spectrum tend to find it useful — a structural form of self-preservation
• The team shipped every week throughout the Twitter-to-X transition; visible execution protected the project from being cut
• Each expansion (pilot → US-wide → global) was proposed only after internal data had already convinced the team it would work; they never asked for permission to take a leap they weren't confident about
• Cost savings were never the justification; the argument was always "this is the only approach that solves trust, scale, and speed simultaneously"

Operating lean at X

• Deleting code matters more than writing it; maintenance debt from small incremental wins accumulates invisibly
• Shared ownership culture: any engineer will jump into any system; teams hand off docs and access rather than guard territory
• Shrinking to people who opted in created an owner mentality that a large assigned team does not naturally develop
• Elon's "Twitter 2.0 / fork in the road" email applied the same self-selection principle at company scale

What's next

• Speed and coverage: ongoing algorithm improvements; expanded BAT-signal feature lets anyone request a note and attach a source to help writers act faster
• Media matching: a single note on an image or video automatically applies to all posts sharing that asset
• LLM-assisted notes (Super Notes): an externally developed research project that uses existing proposed notes as input, generates variants with an LLM, then runs a simulated jury of contributors to predict which variant would be rated helpful — keeping human judgment in the loop while scaling output
• The long-term goal is a product where the algorithm itself is substantially written by the public, not just the notes