Update, October 2026 — Community Notes launched as described and has since expanded across Facebook, Instagram and Threads. The weakness this article predicted — that on polarised stories agreement never forms and no note appears — is now measured rather than forecast. What the numbers show, and how the algorithm produces that result by design, follows the original article.
In January Meta announced it would stop paying independent fact-checkers to label false posts, calling the programme it started in 2016 a failure of trust. On 18 March the replacement goes live: Community Notes on Facebook, Instagram and Threads, beginning as a test in the United States. The idea is borrowed openly from X, where it has run since 2022: ordinary users write short notes adding context to a post, other users rate the notes, and a note is shown only when people who usually disagree with each other both rate it helpful.
How the first phase works
- Around 200,000 people have applied to be contributors across the three apps. They will be admitted gradually, not all at once.
- Contributors can write notes and rate others’ notes, but at first no notes will be visible to the public. Meta says it wants to tune the rating algorithm before anyone sees a result.
- Launch languages are English, Spanish, Chinese, Vietnamese, French and Portuguese. Bangla is not included, and Meta has given no timetable for other languages or countries.
- A note does not remove a post or reduce its reach in this phase. It adds context underneath.
Meta’s argument
“When we launched our independent fact-checking programme in 2016, we were very clear that we didn’t want to be the arbiters of truth,” the company wrote. Experts, it said, “have their own biases and perspectives”, and in the US the programme had not worked as intended. Community Notes is presented as fairer because no single organisation decides; the algorithm only surfaces notes with cross-partisan agreement.
The case against
The same design has a known weakness on X: on fast-moving or highly polarised stories, agreement never forms and no note appears — exactly when one is most needed. Researchers have also found that the majority of proposed notes on X are never shown. Meta is inheriting both problems.
What it means for Bangladesh
For now, nothing changes here. The third-party fact-checking partners Meta uses in Bangladesh continue to work, since the shutdown of that programme is US-only for the moment. But the direction is clear, and Bangladesh — where Facebook is the main channel for political rumour, communal incitement and financial scams, and where Bangla-language moderation has always been thin — has the most to lose if fact-checkers are eventually replaced by a system that does not yet speak the language. The question to put to Meta is not whether Community Notes is a good idea, but when it will exist in Bangla and who will be allowed to write it.
Source: The Verge, via Prothom Alo.
What the numbers show
The system went live on schedule and has expanded across the three apps. The early public figures are modest: more than 70,000 contributors produced around 15,000 notes, of which roughly 6 per cent were published.
That ratio is the story, and it is not an implementation failure. It is what the design does.
Research on X, where the system has run since 2022 and there is more history to examine, finds the same pattern at a larger scale. The Center for Countering Digital Hate reported that 73 per cent of notes on political topics were never displayed, including notes that provided accurate context. The Center for Democracy and Technology found that more than 70 per cent of accurate notes on US election misinformation never reached users, and raised two further problems Meta inherits directly: the model is built for text, and it has no obvious answer for a short video on Reels or for a claim circulating inside a private Facebook group where no public note can attach.
Why the algorithm works this way
The mechanism is easy to misread, because "users rate the notes" makes it sound like a vote, and it is not one. A majority vote would simply let the larger faction label the smaller faction's posts.
Instead the system uses what is usually called a bridging algorithm. From everyone's past ratings it infers a position for each rater on a latent axis — in practice, roughly, which side they tend to be on. A note is then published only if it is rated helpful by people who sit on opposite sides of that axis. Agreement among people who agree about everything else counts for very little; agreement across the divide counts for a great deal.
That is a genuinely clever answer to the problem of who gets to be the referee, and it has a consequence that cannot be engineered away. A note only appears when the correction is not itself contested. For a straightforward fake — a doctored photograph, a miscaptioned video, an invented quotation — cross-partisan agreement is easy and the system works well. For any claim where the facts have become attached to an identity, agreement never forms, nothing is published, and the post stands unannotated.
The failure is therefore concentrated exactly where the harm is: the content most worth correcting is the content least likely to be corrected. And it compounds with a second limit. Notes need time — to be written, rated, and rated by enough people on both sides. Misinformation does most of its travelling in the first hours. Even a note that eventually publishes typically arrives after the audience has moved on.
What it replaced, and what the trade actually is
Professional fact-checking had real weaknesses: it was slow, it covered a tiny fraction of posts, it was expensive, and the question of who appointed the checkers was never answered to anyone's satisfaction. Meta's stated objection — that experts have their own biases and the company did not want to be the arbiter of truth — is not unreasonable on its face.
But the two systems fail differently, and the difference matters. A fact-checking programme fails by being too slow and too small. A bridging algorithm fails by being structurally silent on polarised claims. One of those leaves a gap that more resources would narrow. The other leaves a gap that more contributors will not touch, because the obstacle is agreement rather than capacity.
The language question is unchanged
The point this article ended on has not moved. Community Notes launched in English, Spanish, Chinese, Vietnamese, French and Portuguese, and Bangla was not among them.
The dependency is worth spelling out. A bridging algorithm needs a large, politically varied pool of raters in each language before it can infer the axis it depends on, and needs it continuously, since the axis shifts. A language with few contributors cannot produce published notes at all — not because the software is unavailable but because the statistics do not exist. This is the opposite of professional fact-checking, which scales by hiring: twenty Bangla fact-checkers is a budget line, whereas a Bangla rater pool large and varied enough to bridge is not something a company can simply purchase.
So for markets where a platform carries political rumour, communal incitement and financial fraud in a language with thin moderation, the question to put is still the one from 2025, now with evidence attached: not whether Community Notes is a good idea, but what covers those languages during the years in which the rater pool does not exist — and whether the fact-checking programme is being withdrawn before the replacement can work there.
