Facebook Defends Its Preventive Measures Leading Up to the Capitol Attack

Facebook found itself on the defense over its handling of groups prior to, during and after the violent attacks on the U.S. Capitol Jan. 6 during a press call to detail the release of its Community Standards Enforcement Report for the fourth quarter of 2020. Chief operating officer Sheryl Sandberg said in a livestreamed interview during the Reuters Next virtual summit days after the attack, “I think these events were largely organized on platforms that don't have our abilities to stop hate and don't have our standards and don't have our transparency,” but multiple published reports claimed otherwise, with much of the activity occurring within private groups on the platform. Facebook vice president of content policy Monica Bickert said during the press call that she was “appalled” by the violence at the Capitol, adding that the social network had a team working before the riots to reduce the chances of violent actors organizing via the platform. Bickert said over 890 militarized social movements have been removed to date, encompassing more than 3,000 pages, 19,000 groups and 100 events, and any calls for people to bring weapons to locations were also removed. She also pointed to the social network’s actions against content related to the QAnon conspiracy theory, saying that over 3,000 pages, 10,000 groups and 500 events were removed. As for “Stop the Steal,” which became a rallying cry of sorts for supporters of former President Donald Trump who believed his baseless claims of election fraud, Bickert said public group Stop the Steal 2020, which swelled to over 300,000 members in the two days following the election, was removed Nov. 5, and it began removing all content on Facebook and Instagram containing that phrase prior to Inauguration Day, as the term and related terms were being used to glorify violence. Vp of integrity Guy Rosen said during the press call that content from groups was included in the data the social network shared Thursday, and he offered more color on how groups are handled, adding that over 1 million of them were taken down for violating Facebook’s policies in 2020. “The most important thing to understand for us is that we need to differentiate between groups with some bad actors in them and groups where administrators may be using the group to promote bad behaviors,” Rosen said. “If an admin consistently approves posts that are later taken down for violating our policies, those groups will be taken down,” he added. “Admins must approve all posts in groups about sensitive topics during sensitive times, such as Covid-19. When they don’t, those groups may come down.” He also clarified that Facebook is no longer recommending groups that are civic or political in nature. As for the fourth-quarter-2020 data, Facebook said prevalence of hate speech on its platform fell to 0.07% to 0.08% in the fourth quarter from 0.1% to 0.11% in the third quarter of 2020, and similar trends were seen for violent and graphic content and adult nudity. A prevalence figure of 0.07% means that for every 10,000 views of content, no more than seven violated policy, and Rosen clarified, “Prevalence captures not what we took down, but what we missed.” Facebook attributed its progress to changes that it implemented to reduce problematic content in News Feed, ranking each post to take into account a combination of integrity signals, such as how likely that content is to violate policy, as well as signals from users, including surveys or actions they take on-platform, like hiding or reporting posts. The company also touted improvements in its proactive rate, or the percentage of content that it took action on before it was reported by a user, saying as an example that for bullying and harassment, its proactive rate increased to 49% in the fourth quarter from 26% in the third quarter on Facebook, and to 80% from 55% on Instagram. Facebook credited improvements to its artificial intelligence systems regarding the nuance and context of content, and Mike Schroepfer elaborated during the press call. “There wasn’t one single breakthrough here,” he said. “This was a whole package of AI technologies getting better.” He said mixed media, such as memes with text overlaid on videos or images, are “very difficult for computers to understand,” adding, “From a technical perspective, video is obviously a lot harder to classify than a single photo.” Schroepfer touted his team’s progress in a blog post Thursday, writing, “In the final three months of 2020, we did better than ever before to proactively detect hate speech and bullying and harassment content—97% of hate speech taken down from Facebook was spotted by our automated systems before any human flagged it, up from 94% in the previous quarter and 80.5% in late 2019. More important, this is up from 24% in late 2017—a pace of improvement that we rarely see in technologies being deployed on such a scale.” He also noted that Facebook’s systems now operate in multiple languages, such as Arabic and Spanish, but he cautioned, “We're slowly continuing to regain our content review workforce globally, although we anticipate that our ability to review content will be impacted by Covid-19 until a vaccine is widely available. With limited capacity, we prioritize the most harmful content for our teams to review, such as suicide and self-injury content.” Finally, Schroepfer said Facebook will begin the process of undertaking an independent, third-party audit of its content moderation systems this year. Highlights of the data shared in Facebook’s Community Standards Enforcement Report follow, by category: Adult nudity and sexual activity: Prevalence of this type of content was 0.03% to 0.04% on Facebook in the fourth quarter, down from the third quarter, and action was taken on 28 million pieces of content, down from 36.7 million. Specific results were not provided for Instagram, but its fourth-quarter results were similar to the previous period. Facebook credited an adjustment to its proactive detection technology for photos and videos.Bullying and harassment: Facebook took action on 6.3 million pieces of content in the fourth quarter, up from 3.5 million in the third quarter, and its proactive detection rate rose to 48.8% from 26.4%. Instagram took action on 5 million pieces of content, up from 2.6 million.Child nudity and sexual exploitation of children: Facebook took action on 5.4 million pieces of content in the fourth quarter, down sharply from 12.4 million in the previous period, which the company attributed to a technology adjustment in the third quarter that turned up a large amount of older content that violated its policies. Instagram took action on 800,000 pieces of content, down from1 million. The prevalence rate on both platforms was 0.05%.Dangerous organizations (terrorism and organized hate): Facebook took action on 8.6 million pieces of terrorism content in the fourth quarter, down from 9.7 million in the third quarter, and on 6.4 million pieces of organized hate content, up from 4 million. Meanwhile, Instagram took action on 345,100 pieces of terrorism content, up from 183,200, and on 308,000 pieces of organized hate content, up from 224,700. The prevalence rates were 0.07% on Facebook and 0.05% on Instagram.Fake accounts: Facebook said the total number of accounts it took action on and its proactive rate were similar in the third and fourth quarter, and fake accounts represented roughly 5% of monthly active users in the fourth quarter.Hate speech: Facebook took action on 26.9 million pieces of content in the fourth quarter versus 22.1 million in the third quarter, and its prevalence rate of between 0.07% and 0.08% was down from the third quarter. The company credited improvements in its proactive detection technologies for Arabic and Spanish, along with the addition of Portuguese. Instagram’s content actioned and prevalence numbers changed little between the third and fourth quarters.Regulated goods (drugs and firearms): Facebook took action on 4.3 million pieces of drug content in the fourth quarter, down from 4.7 million in the third quarter, and 1.3 million pieces of firearms content, up from 1.1 million. Instagram took action on 1.4 million pieces of drug content, up from 1.2 million, and on 70,200 pieces of firearms content, up from 36,500. The prevalence level on both platforms was 0.05%.Spam: Facebook took action on 1 billion pieces of content in the fourth quarter, down from 1.9 billion in the previous period.Suicide and self-injury: Facebook took action on 2.5 million pieces of content in the fourth quarter, up from 1.3 million in the third quarter. Instagram took action on 3.4 million in the fourth quarter versus 1.3 million in the third quarter. Instagram also restored 276,500 pieces of content in the fourth quarter, up from 48,000 in the prior period, due to a temporary issue with its proactive detection technology in October. The prevalence level on both platforms was 0.05%.Violent and graphic content: Facebook took action on 16 million pieces of content in the fourth quarter, down from 19.2 million in the third quarter, while Instagram did so on 5.6 million versus 4.1 million. The prevalence level on Facebook was 0.05%.

Politics

Facebook Defends Its Preventive Measures Leading Up to the Capitol Attack

The social network released its Community Standards Enforcement Report for the fourth quarter of 2020

David Cohen

About

Subscriptions

Events

Publications

WORK SMARTER - LEARN, GROW AND BE INSPIRED.

Subscribe today!

David Cohen