Community Notes

Keith Coleman joined Twitter in December 2016, during the company's turnaround under Jack Dorsey. The 2016 election had made the platform's information quality problem visceral—every day brought viral claims, debate threads, and obvious misinformation. Though Coleman spent his first three years focused on user growth and company revival, he kept returning to a nagging question: how would the internet ever solve the problem of misleading information at scale?

By the time he was managing a large PM team, Coleman had become convinced that nothing the industry had tried was working. Fact-checkers couldn't keep pace. Internal trust and safety teams making editorial decisions faced distrust. The problem was unsolved and urgent. So Coleman made an unusual career move: he walked away from managing a large team and asked his boss, K-Von, if he could stop his job and work on "crazy ideas" to tackle misinformation instead. K-Von agreed immediately.

Coleman started with research, consuming everything on the problem and existing solutions. The initial insight was structural: instead of experts deciding what's true, what if the crowd could? But not just any crowd vote—the system needed to find notes that achieved surprising agreement between people who normally disagree on everything.

The original team was remarkably small: one ML engineer (Jay Baxter), front-end engineer, back-end engineer, designer, and researcher. They operated under a "thermal" structure—a Twitter program for isolated, autonomous teams with one clear decision-maker (initially K-Von, later Elon). This structure protected the project from bureaucracy and gave them permission to iterate at startup speed.

Jay Baxter's first algorithm was a PageRank variant focused on anti-manipulation. But it failed at the core challenge: if more users existed on one side of an issue, the algorithm would amplify bias. After this failed prototype, Baxter designed the "bridging-based" algorithm—the real breakthrough. It scores notes based on agreement from people with opposing political views. The insight: when polarized people agree a note is helpful, it's almost certainly accurate and neutral.

The team tested this via internal competition (described as a "Kaggle-style bake-off"), solidifying the approach before production launch.

Community Notes launched as "Birdwatch" in pilot form in 2021, initially limiting contributors to a small group to prove the concept. The team was nervous: could regular people actually do this task? Should they start with journalists or subject-matter experts instead?

They chose the opposite: open it to everyone from day one. The only gate was a verified phone number. Randomized selection of raters ensured fairness and trustworthiness. This openness—"we want all of humanity to participate"—became philosophical bedrock.

When Elon Musk acquired Twitter in late 2022, he noticed the product and loved it. He DMed Coleman saying, "this community notes thing is awesome" and suggested renaming it. Coleman revealed he'd called it "Community Notes" in his very first Figma mockup before the "Birdwatch" pivot. Elon agreed to revert the name. The change was made the next day.

The algorithm's success surprised even the team. External research by UC Berkeley and others showed that seeing a post with a Community Note changed people's agreement with false claims—a result previous literature suggested was impossible.

The impact on virality was staggering: a 30-40% drop in engagement (likes, reposts) in A/B tests, and a network-effect-adjusted 50-60% drop in total spread after notes appeared. One viral misinformation post would tank in reshares once context arrived. Authors deleted their own posts 80% more frequently after being noted.

By setting a conservative threshold (showing only ~8% of proposed notes), the team maintained extraordinary quality. They never ship a note they're uncertain about—the worst failure mode is a bad note that damages trust.

Contributor growth exploded: from a small pilot to 950,000+ contributors nearing 1 million by 2024, all volunteers motivated by intrinsic impact.

In 2024, Community Notes showed 95,000 notes with 30 billion views—more than double the prior year's 37,000 notes and 14 billion views. Coverage spans politics, sports, gaming, entertainment, and breaking news. Notes can be matched to identical images across multiple posts, multiplying impact.

Meta adopted Community Notes as their primary fact-checking tool, validating the approach at scale. The product has survived four different Twitter/X leaders (Jack Dorsey, Parag Agrawal, Elon Musk) because its success is data-driven and undeniable.

Coleman credits the thermal structure and small team with this resilience. With one clear external decision-maker (now Elon), 100% team focus, and freedom from standard planning cycles (no OKRs), the team iterates faster than bureaucratic processes allow. When a goal is set, the entire team focuses on that milestone—whether it takes two weeks or three months—then reassesses.

The most striking testimony comes from an early impact: when a note appeared on a White House tweet, the administration deleted and reissued their statement. A volunteer with 12 followers had shaped national discourse. That's why they keep volunteering.