In January 2019 YouTube started adopting the system. It was at this point that Mark Sargent noticed that his views of the flat earth were skipping. Other types of content have also been downgraded, such as moon landing conspiracies or videos lingering on chemtrails. Over the next several months, Goodrow and Rohe made more than 30 improvements to the system that they say have increased accuracy. By the summer, YouTube publicly announced its success: the playback time of borderline content that came from recommendations was reduced by 50 percent. A drop of 70 percent was reported by December.
The company will not publish its internal data, so it is impossible to confirm the accuracy of its information. However, there are several external signs that the system has had an impact. For one thing, consumers and creators of borderline products complain that their favorite material is seldom improved. “Wow, has anyone else noticed how difficult it is to find ̵
Then there is academic research. Berkeley professor Hany Farid and his team found that the frequency with which YouTube recommended conspiracy videos dropped significantly in early 2019, right at the start of YouTube’s update. According to his analysis, by early 2020, those recommendations had dropped 40 percent from a 2018 high. Farid noted that some canals were not just reduced; they almost disappeared from the recommendations. In fact, before switching from YouTube, he had found that 10 channels – including that of David Icke, the British writer who argues that reptiles walk among us – accounted for 20 percent of all conspiracy recommendations (as Farid defines them). He then found that the recommendations for these websites “basically went to zero”.
Another study, somewhat backing up YouTube’s claims, was done by computer scientists Mark Ledwich and Anna Zaitsev, a postdoctoral fellow and lecturer at Berkeley. They analyzed YouTube recommendations, looked specifically at 816 political channels and categorized them into different ideological groups such as “Partisan Left”, “Libertarian” and “White Identitarian”. They found that YouTube recommendations tend to take political viewers mainstream. The channels they grouped under “Social Justice” on the far left lost a third of their traffic to mainstream sources like CNN. Conspiracy channels and most of the reactionary right – like White Identitarian and Religious Conservative – saw the bulk of their traffic going to commercial right-wing channels, with Fox News being the biggest beneficiary.
If Zaitsev and Ledwich’s analysis of YouTube’s “mainstreaming” traffic holds – and this is certainly a direction YouTube itself advocates – this would fit into a historical pattern. As law professor Tim Wu stated in his book The main switchNew media tend to start in a wild west, then clean up, put on a suit, and consolidate into a prudent center. Radio, for example, started out as a chaos of small operators proud to say something and then gradually got caught up in a small number of mammoth networks mainly aimed at pleasing the mainstream.
However, for critics like Farid, YouTube hasn’t gone far enough fast enough. “It’s a shame on YouTube,” he told me. “After how many years of this nonsense did you finally answer? After the public pressure became so great, they couldn’t handle it. “
Even the executives who put the new “cut back” system in place told me it wasn’t perfect. What makes some critics wonder: Why not just shut down the recommendation system completely? Micah Schaffer, the former YouTube employee, says, “If you can’t do this responsibly, at some point you don’t have to do it anymore.” As another former YouTube employee noted, determined developers can play well any system YouTube puts in place, such as “the Velociraptor and the Fence”.
Even so, the system seemed to work for the most part. It was a real, if modest, improvement. But then the locks opened again. As winter 2020 turned into pandemic spring, summer of activism, and another norm-shaking election season, it looked like the recommendation engine might be YouTube’s slightest problem.