4/10/22, 10:47 Oversight Board | Independent Judgment. Transparency. Legitimacy. disappearance of ethnic Serb civilians. In the video, the city is named Knin and the rats flee on tractors, both references to Operation Storm. Comments to the post make clear these references are apparent to ethnic Serbs and Croatians. The video may contribute to a climate where people feel justified in attacking ethnic Serbs. The post is designed to remind people of past conflict and to rekindle ethnic strife, with the goal of ridding the Knin area of the small remaining Serbian ethnic minority (on historical revisionism and radical nationalism in Croatia see Council of Europe Advisory Committee on the Framework Convention for the Protection of National Minorities’ 2021 Fifth Opinion on Croatia, on online hate speech against ethnic Serbs see the 2018 European Commission against Racism and Intolerance report, para 30). When a threat is “veiled,” according to the Facebook policy, it requires “additional context to enforce.” That context is present in this post. The Board disagrees with Meta’s assessment that the content did not contain threats or statements of intent to commit violence, and that calls for exclusion or expulsion without specifying means of violence may not constitute a violent threat. The forced expulsion of people is an act of violence. The use of the Pied Piper story is not advocacy of peaceful removal but a clear reference to known historical incidents of violence, in particular with the imagery of the tractor. As evidenced by the users who reported this post and the public comments, in the eyes of observers, rats in the cartoon represent the ethnic Serb population of the Knin area, including those who remained there. The cartoon clearly celebrates their removal. In the context of the Pied Piper story, the rats are induced to leave Knin by a magic flute rather than compelled by force, but the tractor reference refers to the actual forcible removal which is widely known about in the country. The tractor is a metaphor, but threats can be conveyed by metaphor no less effectively than by direct statements. Meta’s review process The Board is particularly interested in why the company concluded this content was not violating so many times. It would have been helpful if Meta had focused on this at the outset, instead of focusing on why its revised decision to remove the post was correct. If the company wishes to reduce the level of violating content on its platform, it needs to treat the Board’s selection of enforcement error cases as an opportunity to explore the reasons for its mistakes. The Board notes the complexity of assessing cases such as this one and the difficulty of applying Facebook’s Community Standards while accounting for context, especially considering the volume of content that human reviewers assess each day. Because of these challenges, the Board believes that it is important for Meta to improve its instructions to reviewers and pathways and processes for escalation. “Escalation" means for human reviewers to send a case to Meta’s specialized teams, which then assess the content. According to the rationale provided by Meta, to avoid subjectivity and achieve consistency in enforcement, human reviewers are instructed to apply the letter of the policy and not to evaluate intent. While objectivity and consistency are legitimate goals, the Board is concerned that the instructions provided to reviewers appear to have resulted in about 40 human reviewers erroneously qualifying the content as non-violating, and no reviewer reaching the decision which Meta ultimately believed to be the correct one, which is removal. The possibility to escalate content is supposed to lead to better outcomes in difficult cases. Review on the escalation level may assess intent and is better equipped to account for context. This content was reported 397 times, had a wide reach, raised policy questions, required context to assess and involved content from a Croatian online platform which, according to experts consulted by the Board, was previously the subject of public and parliamentary discussion on the freedom of speech in the context of satire. Yet, no reviewer escalated the content. Meta told the Board it encourages reviewers to escalate “trending content,” and to escalate when in doubt. Meta defined trending content as “anything that is repetitive in nature (…) combined with the type of action associated with the content (i.e. potential harm or community risk(…)).” The fact that the content was not escalated prior to Board selection indicates that escalation pathways are not sufficiently clear and effective. The failure to escalate was a systemic breakdown. One factor which may have prevented the escalation of content in this case is that Meta does not provide at scale reviewers with clear thresholds on when content is “trending.” Another factor that may have contributed to the failure of reviewers to identify the content as “trending” – and thus to escalate – was the automated review system Meta used in this case. Meta explained that it uses automation to respond to reports when there are a certain amount of non-violation decisions over a given time period to avoid re-review. The Board is concerned about Meta’s escalation pathways, and notes that it should provide more information regarding these. It should study whether additional pathways to escalate content are necessary and whether the automated system used in this case prevents content which is https://www.oversightboard.com/decision/FB-JRQ1XP2M 8/11