Mnemonic_Oversight_Board_Comment_2022-002-FB-MR.pdf

How should platforms deal with graphic content that could serve as evidence of human rights violations? That was the central issue addressed in Mnemonic’s recent comment to the Oversight Board. The Oversight Board is the body that reviews content decisions made by Meta on Facebook and Instagram, and in case <a href="https://oversightboard.com/news/1147564895979685-announcing-the-board-s-next-cases/">2022-002-FB-MR</a> the Board is reviewing a decision to remove a graphic video depicting a civilian victim of violence in Sudan. In our comment, we explain how the <a href="https://sudanesearchive.org/">Sudanese Archive</a> archives and verifies content, and asked the Board to “ensure that Meta’s moderation policies and practices don’t lead to loss of such essential documentation.” 
Of note, the comment includes screenshots from a leaked Facebook document titled “Arabic dialect representation in markets” that show that at the time the document was created (October 2020), there were no experts in Sudanese Arabic reviewing content for the platform. In fact, the document notes that “[a]ll highly at-risk countries other than Syria have very low or no representation.” In fact, Mnemonic has reviewed several documents relevant to Meta’s handling of Arabic-language content, and what we have found is deeply concerning. As previously reported by <a href="https://www.politico.com/news/2021/10/25/facebook-moderate-posts-violent-countries-517050">Politico</a> and <a href="https://www.reuters.com/technology/facebook-knew-about-failed-police-abusive-content-globally-documents-2021-10-25/">Reuters</a>, Meta’s handling of content in Arabic is especially deficient. According to a document entitled “An incomplete integrity narrative for MENA”, the platform had failure rates as high as 77% for so-called “terrorist content.”
In these presentations, Meta staff appeal to the platform to hire more dialect experts, to fix poor labeling of content that trains machine learning content moderation processes, and to understand that Arabic truly cannot be treated as a single language. Machine learning already has inherent difficulties with non-Latin characters. Meta’s insufficient investment compounds the problem. We add our voices to these staff in our comment. You can read the full document below: 
<embed src="https://mnemonic.org/uploads/Mnemonic_Oversight_Board_Comment_2022_002_FB_MR_1d2e08d04a.pdf" width="100%" height="800px"/>

BLOG

Oversight Board Case on Graphic Human Rights Documentation from Sudan Has Big Implications

.css-68jq5d{padding-bottom:1.45rem;color:#ffffff;-webkit-transition:0.25s;transition:0.25s;}.css-68jq5d:hover{-webkit-text-decoration:underline;text-decoration:underline;}BLOG

Oversight Board Case on Graphic Human Rights Documentation from Sudan Has Big Implications

BLOG