YouTube鈥檚 recommendation algorithm is the worst at recommending extremist content out of three popular websites. The other two in the comparison were Gab, a social media site similar to Twitter known for being popular with people on the far-right, and Reddit, a news aggregation site.
A team from Swansea University and the University of Stuttgart analysed the way personalisation algorithms on the three websites recommend content to their users.
Three separate YouTube accounts were set up to follow 10 accounts defined as extremely right-wing and 10 that were politically neutral. Each of the accounts then interacted with different types of content: one predominately neutral, one predominately extreme, as defined by a widely-used system called the Extremist Media Index, while the third account did not interact at all.
Advertisement
Each time the account visited the site over a two-week period, the algorithm turned up 18 recommended videos on the YouTube homepage, which were then analysed.
The YouTube account that clicked onto extreme content was recommended extreme content on the home page nearly every time it visited and was twice as likely to be recommended extreme content as the non-interacting account. The neutral-interacting account saw extreme content one in every five sessions.
When a similar analysis was done on both Gab and Reddit, clicks on posts of each category had no impact on what users saw next.
鈥淔undamentally, YouTube鈥檚 algorithm works as it鈥檚 meant to: if you engage with something, it shows you more of it, whether it鈥檚 cooking videos, music or extremist content,鈥� says Joe Whittaker at Swansea University, who presented the work at the聽聽in Swansea last month.
YouTube could instead use algorithms to encourage people to break out of their filter bubbles, says Emillie de Keulenaar at the University of Amsterdam.
YouTube told New 杏吧原创 it can鈥檛 prevent fake or extreme views being published on the site, but it is reducing their prominence.
Partway through the analysis, YouTube announced it would tweak its recommendation algorithms. Changes are still ongoing, but extremist content still exists: only a handful of channels the researchers encountered have been removed.
