An inside look at X’s Community Notes | Keith Coleman & Jay Baxter

Keith Coleman and Jay Baxter explain how Community Notes uses ‘bridging agreement’ to surface accurate context at internet scale, now adopted by Meta.

• In 2024, ~95,000 notes were seen roughly 30 billion times — more than double the prior year’s 37,000 notes seen 14 billion times.
• Only ~8% of proposed notes are actually shown; the bar is deliberately conservative to protect trust in the system.
• Posts with a Community Note see 30–40% drops in likes/reposts in A/B tests and a 50–60% drop in total reposts across the network.
• Authors delete their post 80% more often after receiving a note — a net positive for misinformation, but those high-quality notes then go unseen.
• The algorithm uses matrix factorization to find ‘bridging agreement’ — notes must be rated helpful by people who typically disagree, blocking partisan manipulation.
• Contributors are fully anonymous/pseudonymous by design; pilots showed people cross partisan lines far more readily when not identified by real name.
• During the first three days of the October 2023 Israel-Hamas conflict, 500 notes appeared with a median time-to-publish of 5 hours, vs. 2–4 days for traditional fact-checking.
• The ‘Supernotes’ external research project uses LLMs to generate note variants and simulates contributor ratings to predict helpfulness before human review.

2025-02-27 · Watch on YouTube