X’s Community Notes: A Failing Grade in South Asian Languages
X (formerly Twitter), boasts Community Notes as its flagship crowdsourced fact-checking initiative, designed to combat misinformation through user-generated context. However, a recent study by the Center for the Study of Organized Hate (CSOH) reveals significant shortcomings in the system’s effectiveness, particularly in South Asian languages like Hindi, Urdu, Bengali, Tamil, and Nepali. Despite representing a substantial portion of X’s user base and the global population, these languages constitute a negligible fraction of the Community Notes archive. The study, analyzing 1.85 million public notes, found only 1,608 entries pertaining to South Asian languages, a mere 0.094%. Even more concerning, a mere 37 of these notes ever reached the public timeline.
The crux of the issue lies in the mechanics of Community Notes, which relies on a two-pronged approach: user ratings of “helpfulness” and a “bridging test” requiring consensus among users with differing viewpoints. While theoretically sound, this system falters when there’s insufficient participation from contributors fluent in specific languages. The study demonstrates that South Asian language notes receive comparable, or even higher, “helpfulness” ratings compared to English notes. However, they stumble at the bridging test, with fewer than 40% clearing this hurdle compared to a 65% success rate for other languages. This disparity stems from a scarcity of reviewers fluent in these languages and representing diverse perspectives, hindering the system’s ability to establish the necessary cross-group agreement. Consequently, accurate notes languish in draft form while misinformation proliferates unchecked.
The consequences of this imbalance are particularly acute in South Asia, a region identified as highly vulnerable to misinformation. The underrepresentation of South Asian languages in Community Notes exacerbates existing linguistic inequalities and leaves these communities disproportionately exposed to misleading content. The report further highlights the system’s inability to scale during critical events. Note-writing activity in South Asian languages remained stagnant until the 2024 Indian general election, when a brief surge in submissions overwhelmed the system, leaving a backlog of draft notes precisely when real-time context was most crucial for voters.
Beyond scarcity, the study uncovered a more insidious issue: the weaponization of Community Notes. Even in draft form, some notes revealed troubling content, including ethnic slurs and derogatory remarks targeting specific political groups. This misuse underscores a fundamental misunderstanding of the feature’s purpose and a willingness to exploit it for partisan attacks. The absence of a dedicated moderation layer for Community Notes, relying solely on crowdsourced ratings, further amplifies this risk, particularly in linguistic contexts where hateful speech often carries political weight. Without intervention, Community Notes risks becoming a platform for the very rhetoric it was designed to combat.
To address these critical flaws, the report proposes corrective measures for X, which can also serve as a roadmap for Meta as it develops its own Community Notes system. These recommendations encompass building multilingual reviewer capacity, adjusting publication thresholds to account for linguistic realities, and implementing automated civility filters. Expanding and diversifying the reviewer pool requires sustained outreach in South Asian languages and targeted recruitment drives preceding anticipated events. The algorithm governing note publication must also be calibrated to recognize that smaller language groups will not generate the same vote volumes as English, ensuring that consensus within these groups carries equivalent weight.
The report highlights the urgent need for proactive measures by platforms implementing crowdsourced fact-checking systems. Learning from X’s experience, Meta should prioritize recruitment well in advance of anticipated events, calibrate its algorithms to reflect linguistic disparities, and deploy civility filters to prevent the spread of harmful content. Crucially, success should be measured by coverage parity across languages, not merely by aggregate note counts. Ensuring proportional fact-check protection for all languages is paramount to building trust and fostering truly global communities. Ignoring this imperative risks creating a two-tiered internet, where English content benefits from contextualized information while other languages are left vulnerable to misinformation. In a world increasingly reliant on digital platforms, linguistic equity in trust and safety features is not just a technical challenge but a moral imperative.
