Bridging-Based Agreement
Surface content only when people who normally disagree agree — that's the signal of truth.
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 95%
Instead of majority-rules voting or expert fact-checkers, only publish a note when people who have historically disagreed with each other both rate it helpful. That 'surprising agreement' across a polarized divide is a strong proxy for accuracy, neutrality, and good writing — and it is naturally manipulation-resistant because a voting ring of like-minded accounts can't fake cross-partisan consensus.
Origin
Built by Jay Baxter and the Community Notes (originally Birdwatch) team at Twitter/X starting 2020. Keith Coleman seeded the idea from Chris Bail's research and lists of accounts liked by people on both political sides; related lineage includes Polis, which also looks for agreement among people who typically disagree. Implemented as a matrix-factorization model fit with gradient descent.
Core principles
- 01Consensus across a divide is a better truth signal than raw majority.
- 02You don't need external ground-truth labels or fact-checkers to bootstrap the system.
- 03Bias, not manipulation, is the main failure mode of naive graph/vote algorithms.
- 04Conservative thresholds protect trust — better to show nothing than a bad note.
How to run it
- 1
Let anyone propose context
Any user who has earned the ability can propose a note on any post they think is misleading. Every post is eligible — no exemptions for leaders, advertisers, or the platform owner.
Pro tip Custom-write context for the specific claim and pull it up with zero clicks, rather than linking to a generic 'get the facts' page.
- 2
Collect ratings from a diverse crowd
Other contributors rate the note helpful or not. Deliberately recruit all of humanity so the data captures the full spectrum of viewpoints.
Watch out Filter out raters who consistently rate against the bridging consensus — their ratings stop counting (a reputation component).
- 3
Model the polarized divide
Use matrix factorization to place raters and notes on a latent spectrum, so you can detect whether a note earns agreement from people on opposite sides rather than just one faction.
Watch out A PageRank-style graph algorithm resists voting rings but amplifies the majority side's bias — it does not solve the polarization problem.
- 4
Publish only above a conservative helpfulness threshold
Require a sizable majority on both sides to find the note helpful before showing it to everyone. Tune the bar from real user feedback, keeping it deliberately strict.
Pro tip Only ~8% of proposed notes get shown; hold the bar high because the worst mistake is showing a bad note.
Watch out Also run a separate check for agreement among incorrect ratings so a note can be filtered even if it clears the main threshold.
In the wild
A classic case is an AI-generated or old image reposted as breaking news ('look what's happening in San Francisco' when it's a different city years earlier). A contributor writes a note identifying the real source; cross-partisan raters confirm it.
→ The note shows platform-wide, and because a single note auto-matches every post containing the same image/URL, one note can cover thousands of posts.
The largest deluge of misinformation the team had seen — 500+ out-of-context images in the first days, plus fake battle footage made in the game Arma 3. Critics assumed bridging would find no agreement in such a polarized setting.
→ There were tons of notes; median time from post to note was five hours (vs. two-to-four days for traditional fact-checking). Agreement across the divide was actually easier to find on objectively verifiable facts.
Common mistakes
Using majority-rules voting
A simple majority vote lets the larger faction dominate and produces biased or inaccurate notes; the whole point is to require cross-divide agreement instead.
Assuming you need external fact-checker labels
ML engineers in 2020 assumed you had to keep it closed-source and train on ground-truth labels from fact-checkers. In practice the bridging algorithm bootstraps trust from crowd ratings alone.
Is it for you?
Best for
Product and ML teams building trust, moderation, or ranking systems where neutrality and manipulation-resistance matter more than speed of consensus.
Not ideal for
Cases where no cross-divide agreement exists (deeply entrenched opinions) — the honest output is to show nothing, which won't satisfy teams that need a verdict on every item.
From the transcript
“we actually look for agreement from people who have disagreed in the past”
“when people actually have that sort of surprising agreement that's that's what makes the notes uh so neutral”
“it turns out that you can do that um with with this kind of bridging based agreement algorithm is what we call it”
“if if there's some if there are more users on one side the a pag rank type graph algorithm can actually amplify those biases”
From the episode
An inside look at X’s Community Notes
Keith Coleman (VP of Product) and Jay Baxter (ML Lead)