This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Facebook and WhatsApp were used as primary tools for spreading anti-Rohingya hate speech, dehumanizing propaganda, and fabricated news in Myanmar. This content, often orchestrated by the military and nationalist groups, incited real-world violence, including ethnic cleansing, displacement, and the assassination of prominent figures. Despite repeated warnings from researchers and NGOs, the platform failed to adequately police content or provide sufficient moderation resources in the local language, leading to significant human rights abuses.
Facebook allegedly did not adequately remove anti-Rohingya hate speech, some of which was extremely violent and dehumanizing, on its platform, contributing to the violence faced by Rohingya communities in Myanmar.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Using AI systems to conduct large-scale disinformation campaigns, malicious surveillance, or targeted and sophisticated automated censorship and propaganda, with the aim of manipulating political processes, public opinion, and behavior.
Also classified as
The platform exposed millions of users to extreme hate speech, dehumanizing language, and incitement to violence against Muslims.
Fabricated news and rumors flooded the platform, completely polluting the information ecosystem in Myanmar where Facebook is effectively the internet.
The Myanmar military systematically used Facebook to conduct a coordinated, large-scale disinformation and influence campaign targeting the Rohingya minority to manipulate public opinion and incite ethnic cleansing.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The risk was propagated and amplified by Facebook's automated content recommendation algorithms, which unintentionally facilitated real-world violence while pursuing user engagement goals.
Added categories peak at 4: Severe (Epistemic Harm, Direct)
The systematic exploitation of Meta's recommendation algorithms by the Myanmar military to incite ethnic cleansing represents a critical national security and human rights disaster. It highlights how algorithmic amplification of toxic content can be weaponized by state actors to destabilize society and commit mass atrocities.
The platform's content recommendation and automated moderation algorithms operate autonomously, using learned patterns to rank, distribute, and filter content in real-time without explicit human planning for each decision.
Unacceptable Risk: The reports describe the platform's algorithms being exploited for large-scale behavioral manipulation and influence that directly contributed to real-world violence and ethnic cleansing.
The platform's primary function is to curate and recommend content to users to maximize engagement, while its content moderation systems are designed to filter out violating content like hate speech.
The reports provide extensive, well-documented evidence from multiple credible sources (The New York Times, Reuters, BuzzFeed News) detailing the systematic use of Facebook by the Myanmar military, the platform's technical and operational failures in Burmese-language moderation, and the direct connection to real-world violence and displacement of the Rohingya.
Some might argue that Facebook was merely a passive tool and that the primary drivers were human actors (the military and nationalist groups), as suggested by some government officials in the reports.
Technology
Flawed Automated Detection
AI systems struggled to detect hate speech due to font rendering issues.
Inaccurate Machine Translation
Burmese-to-English translation failed, misinterpreting violent death threats.
Engagement-Driven Algorithms
Recommendation systems prioritized engagement, amplifying incendiary content.
Data Inputs
Coordinated Fake Accounts
Military used inauthentic accounts and celebrity pages to inject propaganda.
Lack of Burmese Training Data
Lack of localized data and linguistic resources hindered AI model training.
Unlabeled Propaganda Content
Lurid photos and fabricated stories were uploaded without metadata flags.
Human Factors
Low Digital Literacy
New users confused Facebook with the internet, believing all shared content.
Active Military Troll Farms
Hundreds of military personnel ran coordinated psychological operations.
Reviewer Language Deficit
Extremely few Burmese-speaking reviewers were employed to evaluate posts.
Process and Methods
Clunky User Reporting System
Platform relied on users to report hate speech via an opaque process.
Outsourced Reviewer Pressure
Reviewers had seconds per post and were trained to keep content online.
No Local Operational Presence
Facebook monitored hate speech from abroad with no offices in Myanmar.
Management
Growth Prioritized Over Safety
Free Basics program expanded user base rapidly without safety safeguards.
Ignored Early Expert Warnings
Management dismissed warnings from activists about ethnic violence since 2013.
Slow Safety Resource Scaling
Company was too slow to hire Burmese speakers and deploy safety resources.
Regulatory Environment
Section 230 Liability Shield
Legal protections shielded the platform from liability for user content.
Ambiguous Local Hate Speech Laws
Vague local laws allowed selective enforcement against government critics.
Lack of Regulatory Oversight
No regulatory bodies forced the platform to police anti-minority speech.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
