Lenny's PodcastKeith Coleman & Jay Baxter: How bridging finds neutral truth
Through bridging-based scoring that rewards agreement between users who disagree; only 7% of proposed notes ever ship, and Meta now copies the algorithm.
CHAPTERS
What Community Notes is—and the core “write, rate, publish” loop
Lenny asks for a clear definition of Community Notes, and Keith explains the basic workflow: users propose notes on potentially misleading posts, others rate them, and only notes found broadly helpful are shown publicly. Jay frames the key differentiator: the system is designed to surface neutral, informative context rather than act like a traditional fact-checking label.
- •Users propose notes on any post they believe is misleading
- •Other contributors rate notes; only high-quality notes get published
- •Goal is adding context, not issuing top-down fact-checks
- •Helpful notes are those that can earn cross-perspective approval
Bridging-based agreement: how the algorithm finds neutrality amid polarization
Jay explains why majority vote or immediate publishing would fail, and introduces the “bridging-based” approach: prioritize agreement from people who have historically disagreed. This surprising-agreement signal helps produce notes that are both accurate and neutrally written, while also improving manipulation resistance.
- •Rejects simple majority vote and instant publishing as too bias-prone
- •Looks for agreement among users with a history of disagreement
- •Surprising agreement correlates with neutrality and accuracy
- •Approach offers anti-manipulation benefits even when open source
Eligibility, contributor reputation, and the philosophy of “all posts, all humanity”
The conversation shifts to who can be noted and who can contribute. Keith emphasizes that every post is eligible (including Elon, governments, advertisers), and contributors must earn note-writing privileges through helpful ratings—while the system still aims to include ‘all of humanity’ to better model what’s broadly useful.
- •Every post can receive notes—no exemptions for powerful accounts
- •Users earn the ability to write notes via rating performance
- •Diverse participation is a feature: ‘all humanity’ improves usefulness
- •Reputation mechanisms reduce the influence of consistently bad raters
Scale and product mechanics: volume, matching, and notifications
Keith shares the scale of the system: hundreds of notes per day, nearly a million contributors, and tens of billions of note views. They describe how notes can match across identical images/videos/URLs to increase coverage, and how notifications reach users who previously engaged with a post once a note appears.
- •Hundreds of notes shown daily vs. far fewer traditional fact-checks
- •Media/URL matching lets one note apply to many duplicate posts
- •2024: ~95k notes seen ~30B times (rapid growth)
- •Users who liked a post can be notified when a note is later added
Publishing threshold and quality control: conservative by design
Keith and Jay describe how notes cross the publication bar, including the (contextual) helpfulness threshold and additional filters for incorrectness. They explain why only a small fraction of proposed notes are shown and why the system is intentionally conservative—trust hinges on note quality.
- •Helpfulness threshold is model-based, not a simple percent of users
- •Additional safeguards filter notes tagged as incorrect despite helpfulness
- •Only ~7–11% of proposed notes are ultimately shown (~8% typical)
- •Quality-first philosophy: the worst failure is publishing a bad note
Behavior change impact: reduced resharing, deletions, and real-world effects
Jay shares research and experiment results showing Community Notes changes beliefs and significantly reduces engagement and reposting on noted content. They discuss network effects (large drops in downstream resharing) and the tradeoff that authors deleting noted posts can reduce the visibility of the best notes.
- •External studies show notes can shift agreement with a post’s claims
- •A/B tests show ~30–40% drops in likes/reposts when notes appear
- •Network-wide repost reductions measured around ~50–60% post-note
- •Authors become more likely to delete posts after being noted
Origin story: Keith’s pivot from management to tackling misinformation at Twitter
Keith recounts joining Twitter in 2016 during the election era, seeing the platform’s influence on public discourse, and later deciding that existing misinformation approaches weren’t scaling or trusted. After paternity leave, he proposed stepping away from his management role to pursue ‘crazy ideas’—which became Birdwatch/Community Notes.
- •2016 Twitter felt like a daily public debate arena
- •Existing solutions (fact-checkers, internal T&S) were slow, untrusted, unscalable
- •Keith chose a hands-on, high-impact project over a traditional management path
- •Early work began with research, prototyping, and iterative validation
Small teams for big impact: the ‘Thermal’ model and operating principles
Keith explains ‘Thermal’ as a structure to protect a dedicated, autonomous team from bureaucracy: one accountable driver, a single senior decision-maker, and full-time focus. They describe why dynamic milestones beat quarterly OKRs and how lean, high-iteration work enabled outsized progress.
- •Thermal teams: isolated autonomy, dedicated staffing, clear ownership
- •Single senior decision-maker enables fast, clean decisions
- •100% focus increases iteration speed and momentum
- •Dynamic milestone-based planning beats heavyweight OKR cycles
Algorithm development: from early prototypes to an internal ‘Kaggle-style’ bake-off
Jay describes early scoring attempts focused on anti-manipulation (e.g., PageRank-like methods) and how real pilot data revealed polarization as the central challenge. That insight triggered a competition-style bake-off to develop better bridging-based scoring approaches grounded in agreement across divides.
- •Initial scoring prioritized manipulation resistance but didn’t solve bias
- •Pilot data revealed polarization as the dominant failure mode to address
- •Internal competition/bake-off accelerated algorithm discovery
- •Bridging-based scoring aligned better with neutrality and usefulness goals
How the team operates day-to-day: the ‘one long Google Doc’ and minimal process
They detail a lightweight operating cadence: daily alignment, a long-running shared doc for coordination, and minimal reliance on heavyweight project management tools. The approach allows rapid reprioritization based on what will most help users right now.
- •Daily team syncs focused on what’s most important and what’s blocking launches
- •A long-running Google Doc serves as the coordination backbone
- •Avoids heavy task management; items can naturally ‘fall off’ when irrelevant
- •Roadmap can change multiple times in two weeks based on new signals
Working with Elon: renaming Birdwatch, lean execution, and ownership culture
Keith shares how Elon endorsed the product, pushed for the clearer name ‘Community Notes,’ and how X’s leaner structure changed execution speed and ownership behaviors. They discuss opt-in culture, faster shipping with fewer layers, and engineering discipline like deleting unnecessary code to reduce maintenance burden.
- •Elon supported the product early and drove the rename to Community Notes
- •Lean teams can scale major features dramatically faster than before
- •Opt-in culture increases alignment and ‘owner’ mindset
- •Deleting code and reducing cruft became essential for maintainability
Launching Birdwatch: proving it works through staged rollouts and expectation-setting
Keith describes a disciplined ‘prove it at every step’ rollout—mockups, cross-spectrum user research, small-scale pilots, and gradual expansion. They considered a ‘dumpster fire’ GIF to set expectations but ultimately focused on clarity, learning quickly that there was “gold” in early notes if they could reliably identify it.
- •Validated concept with prototypes tested across the political spectrum
- •Ran early tests with small pilots before scaling access
- •Early contributor output was mixed, but high-quality notes appeared quickly
- •Core challenge became sifting high-signal notes from the noise
Core principles: people-powered, transparent, no override button, and open source
Keith lays out the foundational principles that made the system trusted: the notes are the voice of the people, not the company; there’s no internal kill-switch for individual notes; and the whole system is transparent with open code and data. Jay explains the real engineering costs of making the system genuinely reproducible externally.
- •Notes represent public judgment, not corporate editorial control
- •No internal button to change a note’s status once it meets criteria
- •Transparency: open-source code and daily published data for auditing
- •Engineering choices were shaped by the goal of external reproducibility
Anonymity/pseudonymity: why hiding identities improved participation and honesty
They explain why the program moved away from real-handle attribution: visibility increased harassment risk and reduced willingness to write on controversial topics. Counterintuitively, pseudonymity increased cross-partisan agreement and honesty, while the system’s quality controls prevented a collapse in note quality.
- •Real-identity attribution reduced participation on controversial topics
- •Pseudonymity increased willingness to cross partisan boundaries
- •Anonymity produced more honest ratings and better bridging agreement
- •Quality mechanisms mitigated typical downsides of anonymous systems
Surviving leadership changes and looking ahead: AI-assisted notes, SuperNotes, and optimism
Keith and Jay reflect on why the project endured across multiple CEOs: it solves a real problem, it consistently proved itself with data, and the output is broadly useful even to disagreeing leaders. They discuss future directions like faster/better notes, improved note requests, and human+LLM collaboration (SuperNotes), ending on a broader optimism about shared truth and societal agreement.
- •Product durability came from measurable outcomes and broad perceived usefulness
- •Future focus: more/better notes, faster publishing, better note requests
- •SuperNotes explores LLM-generated variants plus simulated rating to predict helpfulness
- •Community Notes suggests society has more shared agreement than it appears
