2020-05-22: YouTube's recommended videos get longer as more of them are watched; Most are conspiracy videos.

The video "The NZ Mosque Attack Doesn't Add Up" was recommended from 51 channels

In this post, I examine the results of YouTube's recommendation algorithm through an example of series of videos recommended by YouTube. From this example, I found that:

  • The recommended videos are generated to maximize watch time
  • There is significant correlation between videos' metadata and their recommendation order
  • YouTube's recommended videos promote conspiracy theories (in this example)

Maximizing watch time is YouTube's ultimate goal

YouTube's recommendation algorithm, among other discovery features, focuses on watch time to keep viewers glued to the site. In theory, maximizing engagement benefits YouTube, content creators, and advertisers. It encourages YouTubers to create content that people actually want to watch because it makes them more money from displaying more ads. On the other hand, YouTube makes money from advertisers because they find their YouTube's advertising campaigns responsive so they advertise more. In order to sustain this win-win situation, YouTube must favor engaging videos. Its recommendation algorithm will only suggest such videos to encourage the viewer to stay on the site. Furthermore, YouTube kept track of the content that encourages users to remain on the site in order to reproduce that pattern. This popular method of generating recommendations is called collaborative filtering, in which a video is recommended to a user after a similar user watches it. In other words, if a person A watches videos V1 and V2, then recommend video V2 to person B who watched V1. Collaborative filtering has no way of measuring the truthfulness of a video, therefore, it is possible to recommend content containing violence, hate speech, or extreme conspiracy such as flat earth, staged moon landing, or QAnon videos.

The significant correlation between videos lengths and recommendation order

On 2019-01-23, I watched a series of videos that YouTube recommended to me via its autoplay feature to see how many clicks it will take to go from the first Trending video White House backs down, fully restores Jim Acosta's press pass to a conspiracy video. After 13 clicks, YouTube suggested 97% Owned - Economic Truth documentary - How is Money Created which claims that there is a conspiracy regarding money creation and currency value. After 15 clicks, YouTube started to suggest videos from the well-known far-right extreme conspiracy YouTube channel Infowars. I used AllSides Media Bias Ratings to determine the affiliation of the media outlets and users. For the 19 videos that were suggested, the users are rated as follows:
  1. CNN: Left Leaning
  2. CNN: Left Leaning
  3. CNN: Left Leaning
  4. ABC News: Left Leaning
  5. World News: No Rating
  6. PBS NewsHour: Center
  7. KPIX CBS SF Bay Area: No Rating, but features CBS channels: Left Leaning
  8. ABC10: No Rating, but features ABC News channels: Left Leaning
  9. NBC News: Left Leaning
  10. PBS NewsHour: Center
  11. PBS NewsHour: Center
  12. VICE: Left
  13. VICE: Left
  14. VICE: Left
  15. Independent POV: No Rating (does not feature any popular channels)
  16. The Money Inc: No Rating (does not feature any popular channels)
  17. ChangeDaChannel (affiliated with infowars): Right
  18. ChangeDaChannel (affiliated with infowars): Right
  19. UFOTV On Demand: No Rating (does not feature any popular channels)
  20. FaceLikeTheSun: No Rating (does not feature any popular channels)
I gathered the recommended videos' metadata such as views, likes, length, etc. Analyzing the metadata, I found strong evidence of significant correlation between video length and the order of the video in the series (level of extremism). The recommended videos kept getting longer and longer until it reached an upper limit of three hours and 16 minutes; the average length of the recommended videos after reaching the upper limit is two hours and five minutes. In other words, the recommendation algorithm didn't immediately jump from recommending short videos to long ones. Instead, the length of the recommended videos increased gradually. As of December 2018, the average YouTube video length is 11.7 minutes. Only film and gaming videos averaged above 12.9 minutes.

Correlation between the video length and the level of extremism in the content
There is significant correlation between video length and the order of the video:
Pearson Correlation Coefficient: R = 0.87, P-Value is < 0.00001
Spearman's Rho Correlation Coefficient: Rs = 0.91, P (2-tailed) = 0

No correlation found between the percentage of Upvotes and the level of extremism in the content
No correlation was found between the percentage of Upvotes and the order of the video:
Pearson Correlation Coefficient: R = 0.21, P-Value is 0.38
Spearman's Rho Correlation Coefficient: Rs = 0.16, P (2-tailed) = 0.51

Moderate correlation found between the number of views and the level of extremism in the content
Moderate correlation was found between the number of views and the order of the video:
Pearson Correlation Coefficient: R = 0.63, P-Value is < 0.01
Spearman's Rho Correlation Coefficient: Rs = 0.78, P (2-tailed) = 0.00004

No correlation found between the number of votes/views and the level of extremism in the content
No correlation was found between the number of votes/views and the order of the video:
Pearson Correlation Coefficient: R = - 0.23, P-Value is 0.34
Spearman's Rho Correlation Coefficient: Rs = -0.01, P (2-tailed) = 0.98

I found a moderate correlation between the level of extremism and the number of views. No correlation was found between between the level of extremism and the percentage of up-votes or votes/views.

YouTube's recommended videos promote conspiracy theories

In 2017, YouTube tightened its policies on what content can appear on the platform, or earn revenue for creators, and implemented cutting edge machine learning technologies that scale up the effort of human content moderators. The main target in that change was the bad actors that abuse the platform and spread violence extremism. In addition, combating misinformation became an urgent task for YouTube in 2019 after a long list of research studies, blog posts, and tweets about how its algorithm surfaces videos that spread misinformation such as flat earth, vaccines are made to kill people, 9/11 attacks was an inside job, etc.

YouTube's recommendation algorithm was also changed to surface videos from a wider spectrum because users complained about the poor performance of the content-based filtering in YouTube's recommendation algorithm; they were getting too many similar recommendations, like seeing endless cookie videos after watching just one recipe for snickerdoodles. This could be due to a cold start problem or YouTube's algorithm was tuned for precision too well to recommend videos outside of the scope the user have watched. Although YouTube claimed that this problem has been taken care of, on 2019-02-17, a blogger posted a 20 minute video demonstrating how he got trapped in young girls' videos and accused YouTube of facilitating sexual exploitation of children. He searches on YouTube for "bikini haul" then clicks on the "UpNext" recommended video and soon enough the algorithm started to recommend videos of young girls doing gymnastics, or performing "challenges". Most comments were creepy and obviously were left by pedophiles. He also demonstrated how once a YouTube viewer gets trapped in a rabbit hole, there is no way out. The video went viral and the blogger contacted advertisers like Epic Games Inc, Nestle, and Walt Disney Co who quickly paused advertising spending on YouTube.

Five days later, TeamYouTube tweeted that content creators will be held accountable for the viewers' comments on their videos. "Even if your video is suitable for advertisers, inappropriate comments could result in your video receiving limited or no ads.", TeamYouTube wrote. The company updated its policy regarding how inappropriate content will be handled.

On 2019-02-25 ARS reported that YouTube kids app is hosting cartoons that include spliced-in clips giving children suicide tips. The same clip played in multiple videos and Dr. Free Hess captured a recording of it before she reported it to YouTube and encouraged others to report it. YouTube finally took it down after a few hundred reports.

In an official YouTube blog post, the YouTube team promised to reduce the spread of content that comes close to—but doesn’t quite cross the line of—violating their Community Guidelines. In their words, reducing recommendations of borderline content and content that could misinform users in harmful ways—such as videos promoting a phony miracle cure for a serious illness, claiming the earth is flat, or making blatantly false claims about historic events like 9/11. The software engineer that helped building the AI used to curate recommended YouTube videos, Guillaume Chaslotcalled this change "a historic victory." However, two months later, he tweeted "I was very enthusiast (sic) about YouTube's announcement, but let's be honest: after two months, things barely changed." He posted a link to, AlgoTranperency.org, a tool that shows the videos that have been (are being) recommended by YouTube. On 2019-04-01, shortly after the bloody NZ mosques attacks, I tested the tool and found that the top YouTube's recommended videos are conspiracy videos. One of them was "The NZ mosque Attack Doesn't Add Up - Here's Why" which is a new conspiracy video suggesting that Brenton Tarrant, the attacker, is a hitman and was trained by government intelligence agencies to carryout the deadly attack. The spread of violent domestic extremism is connected to the spread of conspiracy theories according to a local FBI field office.

The video "The NZ Mosque Attack Doesn't Add Up" was recommended from 51 channels
Buzzfeednews reported that YouTube is still suggesting conspiracy videos, hyperpartisan and misogynist videos, pirated videos, and content from hate groups following common news-related searches. It took just nine clicks through YouTube’s “UpNext” recommendations to go from an anodyne PBS clip about the 116th United States Congress to an anti-immigrant video from a designated hate organization. The experiment shows that this behavior is due to YouTube's goal of maximizing users' engagement.

In a tweetPeter Adams, head of education @NewsLitProject, reported that the "UpNext" suggestion algorithm queued up an anti-vaccination video after just one click when he searched "should I get my child vaccinated?" on YouTube (signed out, in a private browser window). Things didn't get better afterwards.

Ben Popken, @NBCNEWS Senior Business Reporter, tweeted about how YouTube's plan to stop surfacing problematic videos has failed. He showed that the top search results for "Antarctica" in a signed-out browser with cleared cache and cookies included a QAnon video, buried aliens, and "bizarre secrets science can't explain." He reported that after one click, YouTube's recommendations are flourished with videos that reject reality, deny science, or contain antisemitism.

No one claims that YouTube's algorithm has been reverse-engineered, but there is undeniable evidence that YouTube's policy of maximizing watch time has been utilized to spread misinformation. Guillaume Chaslot plotted the 84,695 videos that he collected between and 2019-04-18 and 2019-04-25 with the number of views and the number of channels from which it was recommended. Russia Today's take on the Mueller report is the recommendation outlier. It is obvious that manipulating YouTube's recommendation algorithm is not so difficult. Content creators learned what kind of videos play well on YouTube so they try their best to make YouTube's algorithm favor their videos; they make their videos longer and upload lots of conspiracy theories' videos. It is not a fluke that the number of views of an episode questioning the Apollo moon landing in one of RTArabic's shows is 1.3 million views, over 20 times higher than the views of rest of the episodes of the same show. The video title is "Did Americans go to the Moon?" and it is a part of a long playlist of all of the show's episodes. All videos in the playlist are between 25 and 27 minutes except this video. It's a little over 129 minutes.

The problem is that the damage seems to be irreversible or containable. On the NYTimes, Kevin Roose shows how some of the biggest YouTube stars are the ones publishing conspiracy videos, and how their fans are taking them as facts. His example was Shane Dawson, who recently uploaded a 60 minute conspiracy video. His video received over 30 million views although it suggests some of the most ridiculous conspiracies. Among them: that iPhones secretly record their owners’ every utterance; that popular children’s TV shows contain subliminal messages urging children to commit suicide; that the recent string of deadly wildfires in California was set on purpose, either by homeowners looking to collect insurance money or by the military using a type of high-powered laser called a “directed energy weapon.”

A follow-up with over 20 million views claimed that Chuck E. Cheese’s, the restaurant chain, recycles customers’ uneaten pizza slices into new pizzas. The video is so popular that it forced Chuck E. Cheese’s to publicly deny the claim. It is interesting that most of Shane Dawson's videos are not even about conspiracy theories. What does that mean for YouTube? Does that mean that YouTube needs to ban the platform's stars in order to fulfill its promises? Furthermore, is it fair if YouTube punishes its stars for doing what it encouraged them to do to become famous and make a lucrative career out of spreading rumors?

A new independently published paper claimed that, contrary to countless experiments, YouTube’s algorithm has a deradicalizing influence. According to the authors, Mark Ledwich (coder) and Anna Zaitsev (Berkely postdoc), YouTube's recommendation algorithm destroys conspiracy theorists, provocateurs and white identitarians; helps partisans; and hurts almost everyone else. Mark Ledwich's original tweet was posted on 2019-12-28 but the paper was submitted to arXiv on 2019-12-24. Scholars and credible experts dismissed the paper entirely, mainly for not using real users which changes the algorithm's output. Arvind Narayanan, Princeton professor of Computer Science, explained why the paper is not even wrong in a thread of tweets. "The crucial question becomes: what model of user behavior did they use? The answer: they didn’t!" Arvind said. "There’s no good way for external researchers to quantitatively study radicalization. I think YouTube can study it internally, but only in a very limited way." he added. Zeynep Tufekci, professor at the University of North Carolina tweeted a rebuttal to the study “LOL to anyone pretending to study recommendation algorithms without being logged in”. The authors of the paper acknowledged the limitation of their method, but they denied that there is a drastic difference in the behavior of the algorithm between logged-in and anonymous users based on the description of the algorithm provided by YouTube's algorithm developers. Arvind Narayanan added: "I spent about a year studying YouTube radicalization with several students. We dismissed simplistic research designs (like the one in the paper) by about week 2, and realized that the phenomenon results from users/the algorithm/video creators adapting to each other."

The take away message is that YouTube's recommendation algorithm focuses on maximizing watch time regardless of the truthfulness of recommended videos. YouTube is still using recommendation methods that could be manipulated which in turn generate undesirable results. Recommending conspiracy videos by YouTube's algorithm could be a byproduct of maximizing watch-time by recommending longer videos. Conspiracy is the main topic of most recommended videos by YouTube and the damage caused by recommending such videos is, probably, irreversible.

Hussam Hallak