YouTube鈥檚 recommendation algorithm no longer inadvertently sends people down a rabbit hole of extreme political content, researchers have found. Following changes to the algorithm in 2019, individual choice plays a larger role in whether people are exposed to such material.
at the Swiss Federal Institute of Technology in Lausanne and his colleagues used two types of bot to replicate how people interact with YouTube. 鈥淲e came up with this idea of finding a way to disentangle the effects of the algorithm,鈥� says Ribeiro. 鈥淚t feels like finally we came up with a satisfactory answer to a question we started asking long ago.鈥�
One type, the control bot follows the videos watched by one of 125 real-life users, tracking what impact YouTube鈥檚 recommendation system has on their path through the video platform. The other type, a 鈥渃ounterfactual鈥� bot, follows a user鈥檚 behaviour for the first 60 videos they watched, then allows the YouTube algorithm to take control. 聽The counterfactual bots encountered less radical content, as defined by a , than those following the path of the typical user.
Advertisement
The researchers also analysed what happened if a bot watched one or two extreme videos, but then moved back to watching videos from mainstream news outlets. This change was reflected in YouTube鈥檚 recommendations, albeit slowly. For videos recommended in a sidebar while a current video was playing, extreme recommendations disappeared for the counterfactual bots within around 30 videos. However, extreme videos continued to be recommended for a longer period on YouTube鈥檚 homepage 鈥� though the experiment stopped tracking recommendations after 90 videos.
The findings run counter to some previous work on the algorithm, which found that YouTube once suggested extremist content more frequently than Gab, a social network popular with the alt-right. However, it supports conducted after changes to YouTube鈥檚 algorithm made in 2019 by the company behind the platform, which stopped promoting videos YouTube deemed harmful, which it claimed reduced the watch time of such content by 50 to 70 per cent.
In that year, YouTube made more than 30 separate changes to its recommendation system. Today, the company claims that this system now learns from more than 80 billion different parameters, and that its 鈥渧iolative view rate鈥� (VVR) 鈥� what percentage of views on the whole platform come from videos that violate its own policies 鈥� is around 0.1 per cent.聽 YouTube says its VVR has dropped since it started tracking the metric in 2017, but didn鈥檛 say from what level it has fallen.
The paper 鈥減ersuasively challenges serious and well-evidenced allegations over the past six years that YouTube algorithms push, or make more accessible, extreme and far-right content,鈥� says at the Free University of Brussels (VUB), Belgium.
鈥淲e鈥檙e glad to see that the impact of our responsibility work over the years is being noticed. Watch time of authoritative news is up dramatically and borderline viewing is down,鈥� says YouTube spokesperson. 鈥淭his doesn鈥檛 mean we鈥檝e solved all of the issues 鈥� it just means we鈥檒l continue refining and investing in our systems to keep improving.鈥�
Checking whether YouTube does keep improving is important, says Ribeiro 鈥� as shown by the situation before and after 2019. 鈥淭his raises the importance of continuously auditing these platforms and having some sort of continuous study on how things are currently working, because there is very complex interplay between society and technology,鈥� he says.
PNAS
