How X’s Community Notes Leave South Asians Disproportionately Exposed to Misinformation

Community Notes is X’s flagship crowdsourced fact-checking feature, designed to provide timely, user-generated context on misleading posts. In practice, however, the system faces global challenges, including delays, inaccuracies, and inconsistent coverage. These issues are more pronounced in higher-risk languages such as Hindi and Urdu, where accurate notes often stall while misinformation remains visible. A latest study by the Center for the Study of Organized Hate (CSOH), based on a complete archive of 1.85 million public notes, shows that posts in South Asian languages, including Hindi, Urdu, Bengali, Tamil, and Nepali, account for just 1,608 entries, roughly 0.094 percent of the archive, even though the region represents approximately a quarter of the world’s population and five percent of X’s monthly users. Even more striking, only 37 of those notes have ever appeared on the public timeline.

Community Notes relies on two key mechanisms: a “helpfulness” up‑vote and a bridging test, which requires agreement from contributors who typically disagree with one another. The idea is elegant in theory: bring together people with different viewpoints to agree on what’s accurate, avoiding echo chambers. But the system struggles in practice when there isn’t enough contributor activity in a given language to meet its consensus thresholds.

Our findings show that Notes in South Asian languages are marked “helpful” about as often as, and in some cases more often than, English notes. Yet, according to X’s public Community Notes dataset, fewer than 40% make it through the second hurdle, called the bridging test, while roughly 65% notes in other languages succeed. The main problem is reviewer scarcity: there are not enough raters who speak these languages and hold varied viewpoints, so the system cannot confirm the required cross-group agreement. Accurate notes often end up in draft form, while misinformation remains visible.

On X, Community Notes are published at significantly higher rates for English posts than for posts in South Asian languages. Still, they often fail to appear for posts in South Asian languages. This is especially problematic in a region that global bodies, such as the World Economic Forum, have already identified as being most vulnerable to misinformation. This imbalance reinforces existing linguistic fault lines and leaves South Asian audiences with less protection against misleading content.

The report tracks Community Notes authored in South Asian languages from April 2024 to April 2025 and reveals a stark trend. One chart shows note-writing activity remaining nearly flat until the Indian general election window from April to June 2024, when weekly volume briefly surges twentyfold. This pattern tells two stories: first, crises do mobilize contributors; second, the system isn’t designed to scale with that urgency. Drafts pile up precisely when voters most need real-time context.

Weaponized Notes

Scarcity is only part of the problem. Even when Community Notes fail to make it through the system, some drafts reveal troubling intent — for example, calling Indian Congress voters “Italian-mafia supporters” in Hindi or using slurs against Pakistani police in Urdu. That these submissions exist at all highlights both a misunderstanding of the feature’s purpose and a willingness to weaponize it. Because X’s algorithms favor diverse perspectives, a slur or personal insult could theoretically surface if voters from opposing groups happen to share the same animosity toward a target.

Why doesn’t the moderation layer X uses for ordinary posts catch this? Because it doesn’t exist for Community Notes. X relies entirely on crowd-sourced ratings to moderate the tone and language of Community Notes. Yet, those ratings are especially sparse in the very linguistic contexts where harmful speech most often doubles as a political signal. Left unchecked, Community Notes risks becoming a partisan, persistent, and algorithmically amplified channel for the type of rhetoric it was designed to contain.

To move from diagnosis to repair, the report provides corrective steps for X — steps that can also guide Meta as it pilots its own Community Notes system. While the technical implementations may differ across platforms, the design risks are strikingly similar: reviewer scarcity in smaller languages, lack of civility screening, and slow response during election surges. Meta can avoid repeating X’s shortfalls by adopting three measures from the outset: build multilingual reviewer capacity year-round, adjust publication thresholds to reflect linguistic realities, and deploy an automated civility filter at submission. Learning from X’s experience will save Meta time, protect vulnerable users, and prevent the feature from reinforcing the very inequities it aims to fix.

Recommendations for X and a warning for Meta

Our study finds that X’s existing Community Notes system falls short for South Asian languages: accurate notes are stalled, reviewer pools are limited, and misinformation remains visible. On July 1, X announced a pilot that lets AI chatbots draft notes. Automating a feature that already struggles in English and breaks down in Hindi, Urdu, Bengali, and Tamil, risks magnifying the very gaps we documented. Instead, X should focus on fixing the basics before scaling.

It must expand and diversify the reviewer pool to include more speakers of South Asian languages. This will require continuous outreach in those languages, supplemented by more aggressive recruitment drives before predictable events when engagement spikes. The Notes publication algorithm also must recognize linguistic realities. Smaller language groups will not produce the same vote totals as the English language. For example, when a limited but credible group of Hindi, Bengali, or Nepali raters reaches consensus, their agreement should carry weight equivalent to a larger English cohort. This will increase the chances that high-quality notes are published and viewable.

Meta is currently testing its own crowdsourced fact-checking system meant to scale moderation through collective input. A four-day independent snapshot of the US pilot found that only a handful of notes provided meaningful context, while many were inaccurate or irrelevant. Our findings from X demonstrate how quickly such a feature can entrench inequities if its design incentives are left unexamined. Unless Meta confronts these structural gaps, it risks reproducing the same blind spots across Facebook, Instagram, and Threads.

To avoid this, Meta must take proactive steps.

First, it should prioritize recruitment well before moments of crisis. For example, X opened Note authorship in India only weeks before a national election. Platforms must cultivate contributor networks in high-risk languages year-round and conduct targeted recruitment drives six to eight weeks in advance of significant events that are foreseeable.

Second, the algorithm must be calibrated to reflect linguistic realities. Publication thresholds should reflect disparities in the user base. For example, a small number of Bengali raters might provide as much useful feedback as many English raters. Without adjusting the system to account for this, good Notes may go unnoticed.

Third, platforms should implement a basic civility filter at the point of submission. A lightweight language screen can block slurs and other abuse before a draft enters the ratings pool. The marginal latency is negligible compared to the reputational cost of allowing hateful or harmful speech to slip through via a feature designed to protect users.

Addressing these three points as a starting point would give Meta a chance to harness the collective wisdom of its users without perpetuating the very biases a fact-checking system is meant to counter. Critically, the company implementing Notes 2.0 must measure success in terms of coverage parity, not aggregate note count. If Tamil has 80 million native speakers, the platform should be able to determine in real-time whether Tamil tweets receive proportional fact-check protection. Anything less is an illusion of integrity.

Trust and safety for every language

The web’s next billion users are not based in Palo Alto. They represent a diverse global community, speaking languages from every region. If the industry treats multilingual trust and safety as an afterthought, we will have a two-tier internet: content in English will receive context, while everything else will face chaos.

For an industry that prides itself on scale, this is both a market failure and a moral one. More pragmatically, platform executives may soon have no choice. For example, the European Digital Services Act already lets regulators interrogate algorithmic blind spots. Billions of people who speak Bengali, Hindi, Tamil, Urdu, and dozens of other South Asian languages deserve the same line of defense against falsehoods that English speakers already enjoy. To build truly global communities, tech companies must prioritize linguistic equity in product design, as people will judge information quality based on how well all languages are supported.