An inside look at X’s Community Notes | Keith Coleman (VP of Product) and Jay Baxter (ML Lead)

Feb 27, 2025 Episode Page ↗
Overview

This episode features Keith Coleman (Product Lead) and Jay Baxter (Founding ML Engineer) of X's Community Notes. They discuss the product's origin, its unique bridging-based algorithm for combating misinformation, and the principles and lean team structure that enabled its significant global impact.

At a Glance
19 Insights
1h 47m Duration
18 Topics
5 Concepts

Deep Dive Analysis

Introduction to Community Notes

How the Bridging-Based Algorithm Works

Impact and Scale of Community Notes

Note Publishing Threshold and Quality Bar

Philosophies for Broad Participation

Impact of Notes on Content Resharing

Origin Story of Community Notes

Benefits of Small, High-Impact Teams

The Thermal Project Operating Model

Algorithm Development and Internal Bake-Off

Team Operations: Google Docs and Focus

Working with Elon and Lean Operations

Launching Birdwatch and Setting Expectations

Core Principles of Community Notes Success

Anonymity and Pseudonymity for Contributors

Sustaining the Project Through Leadership Changes

Future Directions and AI Integration

Optimism for Society and Community Notes' Role

Community Notes

A system on X where the public can add context to posts that might be misleading. Users propose notes, which are then rated by other contributors, and if found helpful by people who typically disagree, the note is shown to everyone.

Bridging-Based Agreement Algorithm

The core mechanism of Community Notes that identifies helpful and accurate notes by seeking agreement from people who have historically disagreed with each other. This approach helps ensure neutrality and prevents manipulation by polarized groups.

Thermal Project

An internal program at Twitter/X designed to create isolated teams with high autonomy and focus. These teams have a clear owner, a single senior decision-maker, 100% focus on their project, and operate with their own processes, free from typical corporate bureaucracy.

Pseudonymity in Contributions

The practice in Community Notes where contributors can write and rate notes without using their real names or public handles. This encourages more honest feedback and a greater willingness to cross partisan boundaries, as contributors feel less exposed to online harassment or perceived affiliation with a particular side.

Super Notes

A concept for future Community Notes development where existing proposed notes are fed into a large language model (LLM) to generate multiple variants. A simulated jury of contributors then predicts how these AI-generated notes would be rated, aiming to create notes that are highly likely to be found helpful by the community.

?
What is Community Notes?

Community Notes is a system on X that allows the public to add context to potentially misleading posts. Users propose notes, which are then rated by other contributors, and if found helpful by a broad consensus, the note is shown to everyone.

?
How does the Community Notes algorithm actually work?

The algorithm identifies helpful notes by looking for 'bridging-based agreement' – consensus from people who have historically disagreed with each other. This method ensures notes are neutral, accurate, and well-written, rather than relying on simple majority rule.

?
How does Community Notes handle highly polarized contributors or topics?

The system is designed to include all humanity, regardless of their views, to gather diverse data. Notes are only shown if they achieve bridging agreement, meaning a sizable majority of people from both sides of a polarized divide find them helpful, ensuring broad utility rather than catering to extreme views.

?
What is the impact of Community Notes on content engagement and spread?

A/B tests show a 30% to 40% drop in likes and reposts for posts with notes. Capturing the overall network effect, external research groups have found a 50% to 60% drop in total reposts after a note is applied, effectively curbing the virality of misleading content.

?
How many proposed notes actually get published?

Approximately 8% of all proposed notes are ultimately published. This low percentage reflects a very high bar for quality, as the team prioritizes showing only notes that are clearly helpful, neutral, and informative to maintain trust in the system.

?
Why did Community Notes switch to anonymous/pseudonymous contributions?

The switch was made after a pilot revealed that contributors were more willing to write notes on controversial topics and more honest in their ratings when pseudonymous. This increased participation and allowed people to cross partisan boundaries more freely, leading to more accurate and helpful notes.

?
How did the Community Notes project survive multiple leadership changes at Twitter/X?

The project's survival is attributed to the product's inherent ability to produce information helpful to people across different viewpoints, the team's strong execution against ambitious goals, and a consistent focus on proving the product's effectiveness with data at every stage of expansion.

?
What are the core principles that make Community Notes effective and trusted?

Key principles include being the 'voice of the people' (not the company), allowing all humanity to participate, and maintaining complete transparency. The system has no internal 'off' button for notes, and all code and data are open-source and auditable, fostering trust and accountability.

?
How does the Community Notes team operate to achieve high impact with a small size?

The team operates with a 'Thermal Project' model: one clear driver, one senior decision-maker, 100% focus, self-selected members, and a lightweight project management approach using a single Google Doc instead of heavy tools like Jira or Asana.

?
What is the 'Thermal Project' approach to team structure?

The 'Thermal Project' approach involves creating small, isolated teams with a single clear driver, a direct senior decision-maker, 100% focus on their project, and the autonomy to set their own processes and goals, enabling rapid iteration and high impact.

?
How did the Community Notes algorithm get developed?

Initially, a PageRank-like algorithm was prototyped, but data from the pilot revealed it amplified biases. This led to an internal 'bake-off' or competition among ML engineers to develop the bridging-based agreement algorithm, which was found to effectively address polarization and manipulation.

?
What is the future direction for Community Notes?

Future efforts focus on 'more, better, faster notes,' including core product changes like an improved 'bat signal' for note requests. New frontiers involve leveraging AI and LLMs to assist contributors and potentially having the algorithm itself be significantly or entirely built by the public, as seen with 'Super Notes' research.

1. Cultivate Low-Ego Leadership

As a product leader, focus solely on delivering the most helpful outcome for the world, rather than personal power or recognition, which fosters adaptability and trust within the team and project.

2. Empower User-Driven Content

For platforms relying on user-generated content for context or moderation, design systems where the ‘voice of the people’ is paramount, with no company override button, forcing systemic improvements if issues arise.

3. Structure Projects as Thermal Teams

Establish ’thermal teams’ for high-impact projects, featuring a single clear driver/founder, one senior external decision-maker, 100% team focus, and autonomy to define their own process and dynamic goals.

4. Encourage Team Self-Selection

For new projects, allow team members to self-select and opt-in, ensuring they are fully bought into the mission and working style, which significantly boosts success and motivation.

5. Prove Concepts Incrementally with Data

When introducing disruptive ideas or changes within an organization, prove the concept’s viability incrementally with data at each stage, making it harder for others to deem the next step unwise and building internal buy-in.

6. Build Inherently Valuable Products

To ensure a project’s longevity through leadership changes, build a product with inherent, undeniable value that appeals across different viewpoints, set ambitious goals, execute consistently, and always back progress with compelling data and results.

7. Re-evaluate Source of Impact

Don’t equate impact with the number of people managed or the size of your scope; true impact often comes from focused, hands-on building and solving real problems, regardless of traditional career metrics.

8. Build Leaner Companies/Teams

When starting a company or building a team, aim for an even leaner structure than you might initially consider, as small groups can accomplish a surprising amount and thrive due to increased pace of launches and experimentation.

9. Start with Small, Cross-Functional Teams

When initiating a new project, aim for a very small team (e.g., around five people) covering core functions like ML, client engineering, back-end engineering, design, and research to maximize agility and focus.

10. Prioritize Deleting Unnecessary Code

Regularly audit your codebase and prioritize deleting code where the maintenance cost outweighs the incremental gains, especially in lean teams, to prevent long-term maintenance burdens and keep systems efficient.

11. Use Lightweight Project Management

Avoid heavyweight task management tools; instead, use a simple shared document (like a Google Doc) to coordinate, keep the team on the same page, and dynamically manage goals, allowing irrelevant items to naturally fall off.

12. Ensure Full Transparency for Trust

To build trust in a system, especially one dealing with sensitive information, make the underlying code, data, and decision-making processes fully open and auditable, allowing external parties to replicate and verify its operation.

13. Use Pseudonymity for Honest Feedback

When seeking honest feedback or contributions on controversial topics, allow for anonymity or pseudonymity, as people are often more willing to cross partisan boundaries and provide unbiased input without fear of public association.

14. Cultivate Healthy Skepticism

Develop a habit of healthy skepticism towards information encountered online, especially on social media, by recognizing patterns of misinformation and thoughtfully questioning claims to seek a better understanding of reality.

15. Find Optimism in Common Ground

Counter feelings of societal polarization by recognizing that people can agree on a lot, even on controversial topics, and seek approaches that identify and leverage this common ground for progress and collective happiness.

16. Leverage AI for Content Validation

When using AI for content generation (e.g., notes), don’t rely on LLMs to write from scratch; instead, use them to generate variants from existing inputs and then simulate human validation processes to predict helpfulness and ensure quality.

17. Contribute to Community Notes

Sign up to be a Community Notes contributor on X, regardless of your views, as diverse participation helps identify truly helpful and neutral notes, allowing you to have a significant impact on information quality.

18. Propose Notes on Misleading Posts

If you encounter a post on X that you believe is misleading, propose a Community Note to add informative context, as this is the fundamental way the system works to combat misinformation.

19. Set Low Expectations for Launches

When launching a new product or feature, especially one with an unproven concept, be disciplined about setting low expectations and letting the product prove itself incrementally with data, rather than over-promising or expecting immediate perfection.

We actually look for agreement from people who have disagreed in the past. And what we see is when people actually have that sort of surprising agreement, that's what makes the notes so neutral and accurate and well-written really overall.

Keith Coleman

Community Notes adds additional context. It's not fact-checking, necessarily, right? So, there are cases where the post could be true, but maybe it's just misleading because there's no context, or there's missing context.

Jay Baxter

We probably show about 8% of notes that get proposed.

Keith Coleman

The notes just totally take the wind out of these stories, so like the thing will be going viral, note appears, resharing drops 50 to 60%, and like that's it, like it just, you can, at 50 to 60% per generation, the virality quickly goes to zero.

Jay Baxter

Authors become 80% more likely to decrease or sorry to delete their post after they get noted.

Jay Baxter

If there's a problem with a note that's so bad you want to do something about it, it's a problem with the system, like we need to redesign the system to be showing good notes.

Keith Coleman

Society often feels really polarized, you hear people talk about it all the time, like no one can ever agree on anything, but actually like community note shows you people really can agree on quite a lot even on super controversial topics related to politics and everything.

Keith Coleman

Community Notes Contribution Process

Keith Coleman
  1. A user on X sees a post they think is misleading.
  2. The user proposes a note they believe other people might find informative (requires having earned the ability to write notes).
  3. Other people rate that proposed note.
  4. If the note is found helpful by people who normally disagree with each other, it will then show to everyone on X.

Thermal Project Operating Model

Keith Coleman
  1. Identify one clear driver (effectively a founder) for the project.
  2. Establish one clear, senior decision-maker outside the team to go to for decisions.
  3. Ensure everyone on the project is 100% focused on it.
  4. Allow the team to use its own decision-making process, without needing to write OKRs or follow standard corporate practices.
  5. Ensure team members self-select to join the project, aligning their motivation with the mission and team culture.
0.4
Note publishing threshold On an arbitrary scale, indicating the level of helpfulness needed for a note to be shown, set conservatively for high quality.
~8%
Percentage of proposed notes that get published This range has been between 7% and 11% over time, reflecting a high bar for quality.
Hundreds
Notes published per day Compared to approximately 10 traditional fact checks per day.
95,000
Notes seen in the last year (2024) Seen about 30 billion times, more than double the prior year.
37,000
Notes seen in the prior year (2023) Seen 14 billion times.
~950,000
Community Notes contributors worldwide Nearing a million people making the system happen, with more on the waitlist.
30% to 40%
Engagement rate drops (likes and reposts) in A/B test When a post is shown with a note versus without.
50% to 60%
Total reposts drop after a note is applied Measured through a difference-in-differences approach by multiple external research groups, reflecting the network effect.
80% more likely
Author likelihood to delete post after receiving a note This means notes effectively lead to the removal of misleading content, though it also means some good notes are seen less.
5 people
Initial core team size Comprising ML, client, back-end engineering, design, and research functions.
500 GB
RAM required to replicate the algorithm externally To run the Python code on a single machine, demonstrating the transparency principle.
5 hours
Median time for a note to appear during Israel-Hamas conflict From a post going live to a note showing up, significantly faster than typical fact-checking (2-4 days).