How Community Notes reduce viral misinformation | Keith Coleman, Jay Baxter

Episode Summary

Executive Summary: The episode explores X’s Community Notes as a transparent, crowdsourced fact-checking system that relies on disagreement-bridging among users rather than platform authority. Jay Baxter and Keith Coleman explain its open data/code, speed, resilience to manipulation, and growing use with AI-assisted note writing. Audrey Tang frames it as a pro-social model that could evolve from correcting misinformation to surfacing common ground across communities and platforms.

Main Topics: How Community Notes works: A real-world post is annotated by regular users whose notes become visible only after people with differing viewpoints rate them helpful. Notes often provide precise context, correct AI-generated imagery, and clarify misleading but technically true posts. Trust through transparency and no override: The system is designed to be verifiable: anyone can download the code and data, and there is no internal override button. This openness is presented as central to user trust across the political spectrum. Why crowdsourced notes outperform traditional fact-checking: The speakers argue that professional fact-checking was too slow, too small in scale, and too dependent on institutions people mistrust. Community Notes uses polarization itself as a quality-control mechanism by requiring agreement from people who usually disagree. Impact on post visibility and behavior: Posts with notes tend to flatten in views and get reposted less, not because of algorithmic downranking but because users organically engage less once context is added. Authors are also more likely to delete noted posts. Gaming, abuse, and system defenses: The conversation addresses coordinated manipulation, including sophisticated fake-contributor attacks. Defenses include verified phone numbers, similarity detection, reputation signals, random-sample comparisons, and self-correction when bad notes are surfaced. AI-assisted note creation and collaboration: To increase speed and scale, X has opened an API for AI note writers, while keeping human raters in the loop. The longer-term goal is human-AI co-writing, with human feedback improving both note quality and model behavior. From correction to common ground: The discussion ends by imagining a future system that identifies not only misleading content but also ideas and claims that people from different perspectives agree on, potentially creating a pro-social information layer across social media and even politics.

Key Arguments: Crowdsourced moderation can be trusted because the system is open, auditable, and has no internal override, making it 'by the people' rather than by a platform authority. Traditional fact-checking failed at internet scale because it was too slow, too limited in capacity, and lacked legitimacy with users who distrust tech companies. Community Notes works especially well on polarizing topics because people who are predisposed to disagree scrutinize sources more rigorously, producing more accurate notes. A note’s effect is behavioral as much as algorithmic: once users see a correction, they are less likely to repost or endorse the original claim. Manipulation is a real threat, but a combination of identity checks, reputation, similarity detection, and self-correction makes large-scale abuse difficult. AI can meaningfully speed up note drafting, but human judgment must remain the final layer to preserve trust and quality. The same infrastructure that finds consensus on falsehoods could be repurposed to find consensus on constructive ideas, creating a common-ground engine. Open data and open systems allow researchers and other platforms to verify, study, and potentially reuse the mechanism. Reducing misinformation incentives also requires policy changes, such as removing revenue-sharing from noted or undisclosed synthetic-war content.

Data Points: Fact-checking turnaround time: 2 to 4 days - Typical delay for conventional fact-checks, described as too slow for internet time. Community Notes timing: Often on the order of hours - General speed of note appearance compared with traditional fact-checking. Fastest note appearance: As often as about 20 minutes - Possible on brand-new posts when the system can bootstrap quickly. Repost decline after notes: About 50% (2x drop) - Researchers at Stanford, MIT, UW, Paris, and Luxembourg found reposts drop roughly by half after a note is applied. Notes doubled: 2x in the last four months - The team said the number of notes showing on X has doubled recently. AI tools added to a team: 67 - A sponsor ad mention unrelated to the discussion. Security blind spots from AI tools: 67 - A sponsor ad mention unrelated to the discussion. Community Notes contributor scale: Verified phone number requirement - One of the anti-abuse measures for contributors/raters. Older fact-checking scale: 10 or so posts/topics a day - Approximate daily capacity of individual reviewers in previous fact-checking systems. Revenue-sharing suspension for undisclosed AI war footage: 3 months - Penalty for posting AI-generated footage of war/conflict without clearly labeling it. Second offense penalty: Permanent suspension from revenue sharing - If a user repeats undisclosed AI-generated war/conflict footage posting.

Pivotal Quotes: "You can download the real data, the community notes and ratings, run the code on the data to verify that there's no funny business that we're doing on our end" — Jay Baxter / Keith Coleman: Explaining transparency and the absence of an internal override button. "If we can identify common ground at internet scale, it'll make it a lot easier to create a future that humanity likes." — Audrey Tang: Framing the broader pro-social vision beyond misinformation correction. "We have no override button." — Keith Coleman: Answering what happens if powerful people or institutions want a note removed.

Implications: Community Notes suggests transparent, user-governed moderation can outperform centralized fact-checking and may scale with AI. If extended beyond corrections to shared agreements, it could reshape social media incentives toward trust, context, and constructive consensus.