How Community Notes Reduce Viral Misinformation | Keith Coleman, Jay Baxter | TED

Overview

Community Notes is presented as a crowdsourced “context engine” designed to reduce viral misinformation by attaching user-written, evidence-based notes to posts—often without downranking—so people can make better judgments while preserving trust across political divides.

Core mechanism and why it works

• User-written, not official fact-checking: Notes are written by regular users rather than official fact-checkers. Notes are shown broadly only after they become rated “helpful” by people with different viewpoints.
• Credibility via specific, sourced detail: A key example discussed is an AI-generated Iran-related image. The Community Note adds detailed corrections (e.g., explaining why the image doesn’t match reality). The emphasis is that specific, sourced detail is more credible than generic “misinformation” warnings.
• Algorithmic selection that bridges divides: The system is described as using an algorithm that favors “surprising agreement”—selecting notes that are considered helpful by raters who have disagreed historically. This encourages neutrality and reliance on primary sources.
• Eligible for all accounts: Community Notes is framed as applying broadly—even to heads of state and the company itself—meaning there’s no “special treatment” that undermines legitimacy.

Trust and transparency claims

Defenders argue that the system earns trust through openness and verifiability:

• The real algorithm code can be downloaded and audited.
• The platform has no override/veto to remove notes.

The video contrasts this with fact-checking approaches that reportedly failed due to:

• slow turnaround (days),
• limited scale, and
• erosion of trust in centralized systems.

Real-world impact on virality

A described pattern is that posts spread rapidly at first, but once a note appears, views sharply flatten. This is attributed not to algorithmic downranking, but to organic user behavior (people engage less once corrected).

Researchers cited from multiple universities reportedly found:

• substantial drops in reposting (about 50%) after a note is added,
• readers may become less likely to endorse the incorrect core claim.

A tradeoff is noted: because strong notes can reduce ongoing spread quickly, the “best notes” may be seen less often if misinformation cools rapidly.

Addressing manipulation and “gaming” fears

The interview acknowledges that malicious actors will try to game the system, including a hypothetical “Malicious AI Swarm” scenario.

Defenses include:

• verified raters (phone verification),
• signals based on similar historical rating behavior,
• random sampling to detect self-selected or malicious raters,
• rater reputation and other quality controls.

It also emphasizes a self-correcting dynamic: incorrect notes attract rapid attention, but raters mark them not helpful, causing them to stop showing.

Speed and handling fast-changing events

Community Notes is described as typically taking hours rather than days:

• Notes can appear as quickly as ~20 minutes on new posts.
• Notes can appear immediately when matching assets already exist.
• Push notifications are mentioned for users who engage before a note appears.

To accelerate even further, the program opened an AI contribution API:

• AI can draft notes, but humans still rate and correct.
• The model improves over time through human feedback (“reinforcement learning from community feedback”), aiming to produce notes likely to be helpful to disagreeing raters.

Escalating synthetic media and incentive changes

The talk cites a surge in synthetic media tied to the Iran conflict and argues Community Notes is among the highest-scale, highest-speed solutions available, with room to grow (e.g., doubling note volume on X over four months).

It also addresses incentives for misleading posts:

• If posts are noted, creators cannot earn money from revenue-sharing.
• Posting uncited AI-generated war/conflict footage without clear labeling can lead to suspensions from revenue sharing.

Beyond correction: turning polarization into common ground

A major “flipped” vision is introduced: not only show corrections, but also identify content liked by people across perspectives—a “common knowledge engine.”

• A pilot described uses a tag such as “liked by people from different perspectives,” encouraging visibility of shared ground rather than only arguing.
• The proposal extends beyond social platforms to institutions—for example, imagining Congress focusing on issues where agreement already exists to reduce gridlock.

Presenters / contributors

• Keith Coleman
• Jay Baxter
• Audrey Tang
• Divya Siddarth
• TED 2026 curator (Chris) (named only as “Chris” in the subtitles)