Even when users tell YouTube they aren’t interested in certain types of videos, similar recommendations keep coming, a new study by Mozilla found.
Using video recommendations data from more than 20,000 YouTube users, Mozilla researchers found that buttons like “not interested,” “dislike,” “stop recommending channel,” and “remove from watch history” are largely ineffective at preventing similar content from being recommended. Even at their best, these buttons still allow through more than half the recommendations similar to what a user said they weren’t interested in, the report found. At their worst, the buttons barely made a dent in blocking similar videos.
To collect data from real videos and users, Mozilla researchers enlisted volunteers who used the foundation’s RegretsReporter, a browser extension that overlays a general “stop recommending” button to YouTube videos viewed by participants. On the back end, users were randomly assigned a group, so different signals were sent to YouTube each time they clicked the button placed by Mozilla — dislike, not interested, don’t recommend channel, remove from history, and a control group for whom no feedback was sent to the platform.
Using data collected from over 500 million recommended videos, research assistants created over 44,000 pairs of videos — one “rejected” video, plus a video subsequently recommended by YouTube. Researchers then assessed pairs themselves or used machine learning to decide whether the recommendation was too similar to the video a user rejected.
Compared to the baseline control group, sending the “dislike” and “not interested” signals were only “marginally effective” at preventing bad recommendations, preventing 12 percent of 11 percent of bad recommendations, respectively. “Don’t recommend channel” and “remove from history” buttons were slightly more effective — they prevented 43 percent and 29 percent of bad recommendations — but researchers say the tools offered by the platform are still inadequate for steering away unwanted content.
“YouTube should respect the feedback users share about their experience, treating them as meaningful signals about how people want to spend their time on the platform,” researchers write.
YouTube spokesperson Elena Hernandez says these behaviors are intentional because the platform doesn’t try to block all content related to a topic. But Hernandez criticized the report, saying it doesn’t consider how YouTube’s controls are designed.
“Importantly, our controls do not filter out entire topics or viewpoints, as this could have negative effects for viewers, like creating echo chambers,” Hernandez told The Verge. “We welcome academic research on our platform, which is why we recently expanded Data API access through our YouTube Researcher Program. Mozilla’s report doesn’t take into account how our systems actually work, and therefore it’s difficult for us to glean many insights.”
Hernandez says Mozilla’s definition of “similar” fails to consider how YouTube’s recommendation system works. The “not interested” option removes a specific video, and the “don’t recommend channel” button prevents the channel from being recommended in the future, Hernandez says. The company says it doesn’t seek to stop recommendations of all content related to a topic, opinion, or speaker.
Besides YouTube, other platforms like TikTok and Instagram have introduced more and more feedback tools for users to train the algorithm, supposedly, to show them relevant content. But users often complain that even when flagging that they don’t want to see something, similar recommendations persist. It’s not always clear what different controls actually do, Mozilla researcher Becca Ricks says, and platforms aren’t transparent about how feedback is taken into account.
“I think that in the case of YouTube, the platform is balancing user engagement with user satisfaction, which is ultimately a tradeoff between recommending content that leads people to spend more time on the site and content the algorithm thinks people will like,” Ricks told The Verge via email. “The platform has the power to tweak which of these signals get the most weight in its algorithm, but our study suggests that user feedback may not always be the most important one.”
