14 Policy Brief No. 78 Toda Peace Institute 2012, the company did not visit the country until 2014. It first visited on the invitation of the Myanmar government, which invited Facebook to participate in a roundtable discussion following deadly riots in Mandalay which had been triggered by a Facebook rumour. According to Wired, the government sought Facebook’s support during the situation but was not able to make contact and proceeded with shutting down the platform altogether. From this visit onward, the company adopted a bi-annual visit schedule, with the country managed from Australia – a situation which lasted through to 2018, when the company acknowledged that it should do more and started visiting more frequently. Lack of contextual understanding Besides lacking the necessary time to engage with the Myanmar situation, the team in charge of Myanmar lacked the language skills and context to effectively keep up with local developments. Concerns over the company’s lack of understanding of the context and rapidly deteriorating situation were repeatedly raised by Myanmar civil society organisations, including through at least five in-person briefings in Menlo Park beginning in 2014. Despite the repeated warning and urges to hire local expertise, Facebook did not hire Myanmar staff until early 2019. Poor Human Rights due diligence While the UN Guiding principles on Business and Human Rights place a responsibility on companies to uphold human rights, it is unclear to what extent Facebook engaged in the necessary Human Rights due diligence when rolling out its product in Myanmar. The first sign of such an effort being undertaken was the commissioning of an independent Human Rights Impact Assessment, which was published in November 2018, four years after the platform acquired its first million users in the country and more than five years after receiving its first warnings of abuse and weaponisation. Limited use of AI and automated detection While Facebook claims to be leveraging technology to proactively detect community standards violations, it is unclear to what extent the company leveraged automation as part of its Burmese language moderation prior to 2018. As per the company’s own reporting (here and here), only 13% of the content removed in the 4th quarter of 2017 was proactively identified. Judging by the number of reported removals in Myanmar in the 3rd quarter of 2018 (64,000), this would have represented at most 2000-3000 self-identified violations a month, raising questions about the scale and effectiveness of the system. The mass takedown of content containing the keyword ‘kalar’ in May 2017 (a derogatory slur used against Muslims of South Asian descent – but also a composite in harmless words such as kalar pae /chickpea or kalar htine / chair ) also raised widespread concerns over the use of overly simplistic keyword-based detection, failing to account for lin-