4/10/22, 10:45 Oversight Board | Independent Judgment. Transparency. Legitimacy. allowances built into the Dangerous Individuals and Organization policy, Meta clarified that “we require people to clearly indicate their intent. If a user’s intention is ambiguous or unclear, we default to removing content.” Meta also explained that it prefers its reviewers not to infer intent as this “helps to reduce subjectivity, bias, and inequitable enforcement during content review while maintaining the scalability of our policies.” Meta informed the Board that it was unable to explain why two human reviewers incorrectly removed the content and did not properly apply the allowance for reporting. The company noted that moderators are not required to document the reasons for their decision beyond classifying the content as part of their review — in this case, as a violation of Meta’s Dangerous Individuals and Organizations policy on the grounds of praise. In response to the Board questioning whether praise of dangerous organizations can be disseminated as part of news reporting, Meta stated its policy “allows news reporting where a person or persons may praise a designated dangerous individual or entity.” Answering the Board’s question on the difference between standard and severe strikes, Meta explained that the strike system contains two tracks for Community Standards enforcement: one that applies to all violation types (standard), and one that applies to the most egregious violations (severe). Meta states that all violations of the Dangerous Individuals and Organizations policy are treated as severe. The company explained to the Board that severe strikes are those that apply stricter penalties against more serious harms, and limit access to high-risk services such as Facebook Live Video and ads. Meta also referenced a page in its Transparency Centre on “Restricting Accounts” (updated February 11, 2022) that it says explains its approach to strikes. In response to the Board’s questions, Meta provided further explanation of its systems for correcting enforcement errors and how those impacted this case, leading to the Board asking several follow-up questions. The content in this case was automatically detected and sent to a High Impact False Positive Override (called HIPO by Meta) channel. This is a system designed to correct potential false positive mistakes after action is taken on the content. Meta clarified to the Board that this system contrasts with Meta’s General Secondary Review system (part of the cross-check program), which is designed to prevent false positive mistakes before action is taken on the content. Content sent to the HIPO channel joins a queue for additional review, but review will only occur where capacity allows. The position of content in the queue depends on a priority score automatically assigned to the content. Meta explained that content is prioritized for HIPO review based on factors including: topic sensitivity (if a topic is trending or sensitive); false positive probability; predicted reach (the estimated number of views the content might obtain); and entity sensitivity (the identity of the group or user sharing the content). Meta explained that content can be restored in two ways: either the specific piece of content is reviewed by moderators and found to be non-violating; or Meta’s automated systems find that the content matches other content that has been reviewed and determined to be non-violating. The page of the news outlet was previously subject to cross-check, but as part of a platform wide update to the cross-check system, the page was not subject to cross-check when the case content was reviewed, and cross-check did not impact the review of the case content. According to Meta, cross-check now includes two systems: General Secondary Review applies to all organic content on Facebook and Instagram; Early Response Secondary Review applies to all content posted by specific listed entities, including some news outlets. Meta stated that when content from those specific entities is identified as violating a content policy, instead of being enforced, it is sent for additional review. It is first sent to Meta’s Markets team. If a reviewer on this team finds the content is not violating, the process ends, and the content remains on the platform. However, if a reviewer on this team finds the content is violating, it is escalated to another team. This team, the Early Response team, is made up of specialized Meta content reviewers. A reviewer on this team would need to find the content violating before it can be removed. At the time the case content was identified as violating, the news outlet page was not on the Early Response Secondary Review list in the current cross-check system. Additionally, the case content in question was not reviewed as part of the General Secondary Review system, which would also involve additional review before enforcement. According to Meta, the content was sent to the HIPO channel after it was removed, but it was not prioritized for human review. It did not receive an additional human review “due to the capacity allocated to the market” and because the content in this case was not given a priority score by Meta’s automated systems as high as other content in the HIPO queue at that time. Content prioritized by HIPO is only reviewed by outsourced reviewers after an enforcement action is taken. Meta allocates Urdu reviewers to different workflows based on need. These reviewers are shared across multiple review types, meaning they are not solely dedicated to a single workflow. In mid-2022, Meta’s HIPO workflow had less than 50 Urdu reviewers based on that need at the time. The Board asked Meta 34 questions. Meta responded to 30 fully, three partially and declined to answer one. The partial responses were to questions on: providing the percentage of removals under the Dangerous Individuals and Organizations policy that are restored on appeal or second review; accuracy rates for enforcing the prohibitions on praise and support in at-scale review; and how Meta determines intent for the reporting allowance and applicable contextual factors. Meta left one of the questions on providing data regarding the volume of Dangerous Individuals and Organizations content that is removed through automation versus human review unanswered on the grounds that it was unable to verify the requested data in the time available. https://www.oversightboard.com/decision/FB-U2HHA647 6/12