Facebook ranked Myanmar's feed for engagement while it could not moderate Burmese

Facebook was the internet in Myanmar

When Myanmar came online after 2011, it came online almost entirely through one application. Cheap smartphones shipped with Facebook preloaded, and for most of the country's roughly 20 million internet users Facebook was the internet itself.¹ The company ran the same feed there that it ran everywhere: posts ordered not by time but by how much engagement the system predicted each one would draw.

Between the start of the military's clearance operations on 25 August 2017 and the middle of 2018, nearly 725,000 Rohingya fled Rakhine State for Bangladesh.² The UN's Independent International Fact-Finding Mission on Myanmar found the Tatmadaw, the country's military, carried out the operations with genocidal intent. It found enough evidence to prosecute the senior commanders for genocide.² Close to one million Rohingya refugees remain in camps around Cox's Bazar today.³ The military planned and carried out that campaign. This lesson is about the smaller, contested question the record leaves open: what the feed did alongside it.

What an engagement-ranked feed rewards

A feed like Facebook's does not show posts in the order they were written. It ranks them. For each post a person could be shown, the system predicts how likely that person is to react, comment, or share, and it puts the posts with the highest predicted engagement at the top. The aim is attention: a feed that holds attention keeps people scrolling and coming back.

That aim has a reliable side effect. The posts that pull the strongest reactions are not the calmest ones. Content that provokes outrage or fear earns comments and shares at rates ordinary posts do not, so a system that ranks by predicted engagement hands that content more reach than a plain timeline would. Nothing in the ranking picks out hate speech to promote. It promotes whatever people react to, and people react to alarm. With no countermeasure pushing the other way, inflammatory content is amplified as a matter of course. This is not specific to Myanmar or to Facebook. It is how engagement ranking behaves, a pattern documented across Facebook's own internal research, later surfaced in the 2021 Facebook Papers.⁴

None of this is the system breaking. The recommender did what an engagement-ranked feed is built to do. Where that design meets a steady supply of incitement and nothing removes it, the incitement is what travels.

The moderation that could not read Burmese

Amplification is only half of what decides which posts spread. A platform also takes down content that breaks its rules, and at Facebook's scale that removal is its own system. No one can read billions of posts a day by hand. Moderation runs on two things: automated classifiers that flag likely violations for a human to review, and reports from users that send a specific post to a reviewer. Both depend on language. A classifier has to be trained on the language of the post, and a reviewer has to be able to read it.

For Burmese, neither was in place while the incitement spread. Facebook had no working automated detection for hate speech in Burmese during those years, and almost no one to read what users reported. Reuters, working with the Human Rights Center at the University of California, Berkeley, found the company had a single Burmese-speaking reviewer in 2014 and roughly sixty by the middle of 2018, for those 20 million users. It had no employees inside the country, and routed Burmese moderation through a contractor, Accenture, in Kuala Lumpur.⁵ Detection was reactive by default: with no classifier that could read the language, the system waited for a user to report a post before a reviewer in another country ever saw it.⁵

The company's own figures show how late the capacity arrived. Facebook reported 99 Burmese-language reviewers by the end of 2018, and said the share of hate speech its systems caught before any user reported it rose from 13 percent in late 2017 to 63 percent by the third quarter of 2018.⁶ Both improvements came after the August 2017 campaign, not before it.

Even read generously, enforcement stayed thin. An internal Meta document from July 2019, quoted by Amnesty International, put the share of hate speech the company acted on across the whole platform at about 2 percent.⁴ That figure is platform-wide rather than specific to Myanmar, but it describes the ceiling the Burmese effort was built under.

What the record settles and what it cannot

The most careful account of the platform's role is the UN Mission's. It called social media's role "significant" and Facebook "a useful instrument for those seeking to spread hate," and in the same passage declined to assert that the posts caused the violence, saying that link "must be independently and thoroughly examined." It also recorded that Facebook could not give the Mission country-specific data on how hate speech spread there.²

Facebook's own commissioned assessment, written by the consultancy BSR, framed the harm as the work of "bad actors" and "a minority of users." It stated plainly that "the actual relationship between content posted on Facebook and offline harm is not fully understood."¹ Facebook itself went only as far as a moderation failure. Its product policy manager, Alex Warofka, wrote in November 2018 that the company "weren't doing enough to help prevent our platform from being used to foment division and incite offline violence" and that it "can and should do more."⁶ That is an admission of not removing enough, not of an algorithm that promoted the incitement.

Amnesty International went furthest. In 2022 it concluded that Meta's "content-shaping algorithms proactively amplified and promoted" anti-Rohingya incitement, and that the company had "substantially contributed" to the harm rather than being a "passive and neutral platform."⁴ That conclusion rests on the general Facebook Papers evidence about how engagement ranking behaves, plus inference. It does not rest on Myanmar-specific data showing how much anti-Rohingya content the feed actually boosted, because no such data is in the record: Facebook told the UN it could not provide it. "Proactively amplified" is Amnesty's reading, not a measured share.

There is a real counterargument about the mechanism. Writing in Lawfare in 2018, Evelyn Douek noted that the anti-Rohingya campaign was "not merely organic": later reporting showed that roughly 700 military personnel ran a covert operation to seed and spread it. She drew on the scholar Daphne Keller's distinction between harm that comes from a system's "innate structural flaws" and harm that comes from deliberate exploitation by bad actors, which call for different responses.⁷ A feed that amplifies whatever draws reaction is one problem. A military that engineers what draws reaction is another. Here the two are hard to pull apart.

Facebook did act, late. On 28 August 2018, citing the UN Mission, it banned 20 people and organizations. It removed 18 Facebook accounts, one Instagram account, and 52 pages followed by nearly 12 million people, among them the commander-in-chief, Senior General Min Aung Hlaing, and the military's Myawady television network.⁸ In December 2021, Rohingya refugees in the United States sued Meta for at least $150 billion, alleging that its algorithms prioritized "divisive and polarizing content."⁹ Even that complaint pleads Facebook as "a substantial cause," not the cause.⁹

What would settle the size of the platform's contribution is data that does not exist. It would take a record of how the ranking gave reach to anti-Rohingya content against a neutral baseline, and a defensible account of how far the campaign would have spread without the platform. Neither is available. What the record does support is narrower and still serious. Engagement ranking amplified inflammatory content, and near-absent Burmese moderation let it run with little friction. Together they widened the reach of a campaign the military planned and drove. A substantial contribution of contested size, not a measured share and not the primary cause.

Where the same feed runs now

The conditions that made Myanmar dangerous were not unique to Myanmar. The same engagement ranking runs on every large feed, and it rewards the same high-arousal content everywhere. Amnesty's reading of the Facebook Papers described the news feed, ranking, and recommendation features as systems that "actively amplify and distribute content which incites violence," a description of the ranking rather than of one country.⁴ What varied in Myanmar was the second half of the pattern: near-absent moderation in the local language, in a country the company had no presence in, as tens of millions of people came online at once through the platform.

That combination is not history. It recurs wherever a platform's reach into a new market outpaces its ability to read what spreads there. An engagement-ranked feed carries whatever draws reaction. What limits it is moderation the company has to choose to build, in the language the people there actually use.