“problematic content,” the Cornell Tech team concluded.59 Based on analysis of 136,000 channels and 71 million videos, the researchers found that 61% of “alt-right” and misogynist creators are pointing their YouTube followers elsewhere—for example, to buy politically themed products and make direct contributions via fundraising sites such as patreon.com.60 Election falsehoods The extraordinarily bitter 2020 presidential election and its aftermath, combined with disruptions related to Covid-19, have provided opportunities to assess YouTube’s efforts to reduce various forms of harmful misinformation. Like its social media peers, YouTube has stepped up content moderation activities on both fronts—but not enough to prevent a continuing flow of false and conspiratorial content. The Election Integrity Partnership, which brought together academic, civil society, and private sector researchers, published a revealing assessment of how misinformation spread from a variety of sources and was then further amplified on Twitter. YouTube was the third most-linked domain in the partnership’s sample of consequential falsehoods shared on Twitter, behind only far-right sites The Gateway Pundit and Breitbart. Starbird, the University of Washington researcher and one of the partnership’s contributors, described YouTube’s role as “a place for misinformation to hide and be remobilized later.”61 One example mentioned in the partnership’s final report in 2021 was a YouTube livestream by right-wing creator Steven Crowder titled, “Democrats Try to Steal Election!?” The video, which was posted on November 4, 2020, and has tallied more than five million views, remains available. YouTube appended a text panel—one which appeared with all 2020 election-related content— saying, “Robust safeguards help ensure the integrity of election results.” The company also provided a link to the 12 YouTube’s Data Gaps Social media companies collect users’ personal data so that marketers can target specific types of consumers. The platform companies also analyze prodigious amounts of content in an effort to police their platforms for material that violates their rules. But for all of this data analysis, Google, YouTube, and other platforms withhold key information that the public and policy makers need to form sound opinions about the effectiveness of content moderation. Google discloses statistics about the number of YouTube channels and individual videos that are removed for repeated violation of its guidelines on violence, hate speech, harassment, and the like. But it’s difficult to know what to make of these numbers because they represent the numerators of fractions lacking denominators. Consider channel removals. In the final three months of 2021, YouTube took down 3.9 million channels, 90% of which were removed because they trafficked in spam, misleading content, or scams, according to Google’s latest Transparency Report. The next-largest category, with about 4%, were channels featuring sexual content or nudity.1 But how impressive is the removal of 3.9 million channels? It’s hard do say, since Google doesn’t estimate how many rule-breaking channels there were in all. That larger universe would provide the denominator for the fraction, constituting the platform’s success rate. The company reports that of those videos taken down, about one third are removed by its automated screening system before anyone sees them, roughly another third are caught after they’ve been viewed no more than 10 times. The final third come down after having been viewed 11 or more times. So, what can we tell from that? Again, not a lot. Artificial intelligence has improved enough to nab millions of rule-breaking videos before they do any harm. But most of those automated removals probably concern garden variety spam—relatively easy to detect, potentially annoying, but not exactly corrosive of democracy. Of the one-third slice that do get seen, we don’t know whether they are viewed 11 times or 11 million times. Google doesn't say. Another set of missing data are statistics on false positives: how often YouTube’s moderation systems mistakenly remove videos, such as those meant to document human rights abuses or war crimes. In April 2021, Google began releasing a new metric: YouTube’s “violative view rate.” VVR estimates the proportion of video views of content that violates platform rules (excluding spam). In other words, VVR measures how often users encounter harmful content. Google says that VVR is moving in the right direction, which is to say downward, by 81% since the fourth quarter of 2017. 1 2 https://transparencyreport.google.com/youtube-policy/removals?hl=en https://transparencyreport.google.com/youtube-policy/views?hl=en HOW YOUTUBE SPREADS HARMFUL CONTENT – AND WHAT CAN BE DONE ABOUT IT