YouTube’s Black Box of a Recommendation Algorithm

Google made a monster, and it’s having major trouble containing it.

Imagine doing your usual stop at a local coffee shop, only to after minutes of having gotten your usual order of iced coffee with milk, the waiter chimes in unbeknownst, wondering if you’d like to eat raw Ghost pepper based on what you just picked up. On top of it sounding incredibly odd and just creep-inducing on a level most-primitive, you’d immediately start wondering what did upper management do to warrant such a choice so seemingly illogical, to become something they’d even consider doing.

Such is the status of modern YouTube. The platform has made itself a habit of promoting content to users that they may not have any cursory interest in, but being the naturally-curious beings we are, some click on that content and may end up enjoying it–maybe even become influenced by it. If the extent of these recommendations was cooking videos and travel vlogs, the purpose of this story would become moot; but contrary to a much-lauded narrative from YouTube’s public-facing figures, the algorithm does not only stop at suggesting you content within the lane you’re interested in, or even branch you out a little bit to eke out extra watch-time–it is in extreme cases radicalizing individuals by preying on their most primitive propensity for consuming content, or weaponizing their unhealthy relationship with the internet to maximize their revenue-generating potential, without a care in the world for the ethical and moral implications of such an approach to content discovery.

The problem with pinpointing issues in the algorithm is the utter lack of transparency from YouTube on how it exactly works–it is so shrouded in mystery, that ex-Googler Guillaume Chaslot, who worked on the algorithm, had to make the “AlgoTransparency” initiative to track its occasionally-problematic output. Whistleblowers and AI experts in academia can only reveal so much though, and the only entity capable of revealing the full story — and subsequently allowing public scrutiny to help improve its product — is Google. With that lack of insight into what the algorithm does, users are free to make up their own narratives about whom it benefits, versus whom it demotes, but one trend remains a constant in all peripheral analysis of the algorithm–it is very heavily personalized to each user’s watching habits, and that in itself constitutes the biggest difficulty with breaking down how it exactly works. Big trends can be looked at independently, but it’s nothing quite definitively conclusive. Still, there is helpful literature in and outside of YouTube that help construct a semi-cohesive picture of how the algorithm works, and what it aims to do for the company aside from resulting in an upward curve of user engagement.

Derek Muller aka Veritasium helped shed a light on the algorithm’s most-basic function–interest-agnostic content discovery. If you’ve been an active user in the platform for the last few months, chances are you’ve seen a thumbnail of himself standing atop a boat stranded by a ridiculous amount of shade balls in a reservoir. The video in question, is called “Why Are 96,000,000 Black Balls on This Reservoir?” and it boasts north of 35M hits on YouTube. The most interesting part about this video is it doesn’t fall out-of-line for what Derek usually does with his channel — it’s usually breaking down the scientific rationale behind a certain technology or some strange phenomena — but yet, it managed to get dozens of times more views than his other videos, which for all intents and purposes, are almost identical in format and execution.

So, what did cause this to happen in the first place? The first-hand explanation comes from Derek Muller himself, whose channel was experiencing a massive downturn in engagement after Science YouTube had seemingly hit its stride. He goes on to break down a conversation he had with MrBeast — who’s notoriously well-versed on the ways of the algorithm — and he suggested among other things, to target a higher-than-average CTR — clickthrough-rate for the uninitiated — and what that essentially boils down to, is a percentage calculated by way of proportioning the number of times a video has been clicked on, from the amount of times it has been shown to present that opportunity in the first place. What Derek goes on then to assert, is that this will have to force him to pick more easily-marketable topics to cover, as well as shamelessly rely on clickbait-y titles and thumbnails to drive the engagement of his videos in such a way that it can make up for the weeks and months of painstaking preparation that goes into them.

For Derek Muller, and many other creators, opting in for a more commercially viable approach to YouTube is not a matter of moral principle or delivering a sure product to your audience when all currents point the opposite way–it is rather a question of survival, and YouTube has continuously changed the way in which creators have to interact with their analytics in order to maximize viewership, so the introduced barrier to publishing consistently-viewed content on YouTube isn’t production-value, or even brand-association anymore. It’s become YouTube literacy.

That literacy has been somewhat of a complicated web of sometimes-relevant, and sometimes-outdated information whose propagation is impeded by the very nature of content permanency on YouTube. Simply put, if you watch a well-researched video on how the algorithm works made even one year ago, chances are its usefulness will only become rooted in historical value. The cards are constantly being shuffled, and the deck is ever-altering, that learning the unspoken rules of the game has become a sophisticated exercise in gambling, as opposed to any permanent knowledge creators can use so that they don’t have to split their time between making content, and agonizing over what best suits the algorithm days, or even weeks after the content is released.

That YouTube relies on maximizing engagement to make its business-model viable is perfectly-understandable — shy of only suggesting the government should subsidize it given how ubiquitous it has become — but the problem with the algorithm is that it seems like there is some level of deliberation, or what some AI experts would dub as an inherent flaw to its design that makes it so hard to assert that it’ll always make the decisions that are simultaneously in the best interest of YouTube’s bottom-line, but also its users. In light of recent controversies, analysis done by independent parties has suggested the algorithm’s suggestion properties aren’t exactly human-intervention-free, and that has thus allowed some of the biases of its conceivers to seep in, causing the platform to radically shift its profile from a place where people just harmlessly talk about their daily grievances on a poor-quality webcam, to becoming this behemoth of an opinion-shaper where the most valuable currency is that of audience-retention.

Most-recently, a YouTuber who goes by the name of “Coffee Break” broke down a collected dataset of videos on the Trending tab from November 2017 to June 2018 made by Mitchell Jolly, a student of Computing Science at the University of Glasgow. What the conclusion of Mitchell’s findings seems to suggest, is that when channels are categorized in a rough split between traditional media (your CNN’s, Vox’s, and whatnot) and the creators everyone goes to YouTube to watch, the Trending tab seems to skew very heavily in favor of traditional media, especially — and perhaps most-suspiciously — in the US. Stephen chalked it up to some level of human intervention in the US where traditional media is given a lower ceiling to hit to make the Trending tab because those tend to be the safest choices for advertisers, given independent creators have a consistent tendency to be more “provocative” and “edgy” than traditional media.

Stephen’s video tries to shy away from any explicitly-political analysis of this finding, but what this seems most to suggest is that the deliberation at the heart of YouTube’s preferential treatment of traditional media is not some undue patronage to the elite class in America–it’s rather fright of further regulatory oversight, and in most cases recently, an effective method to cushion the blow of potential controversy. Just a few months before this dataset was collected, YouTube was caught in the thick of a major controversy around ads appearing next to extremist content, prompting some of the platform’s biggest contributors of advertisement money to momentarily suspend their ad rolls. This pushed YouTube CEO Susan Wojcicki to pledge at advertisement festival NewFronts just before the study was conducted, that the platform was going to remedy the situation. It’s unclear without that spur of negative attention around the company had anything to do with the way it primarily promoted content on the Trending tab, but its reverberations were felt most-strongly throughout YouTube’s other discovery channels–mainly through the autoplay feature, and the ‘Up next’ column.

This does correlate — though somewhat unintuitively — with a pattern of poor decision-making from the YouTube content moderation team that seems to suggest that they aren’t as much concerned about the nature of extremist content as long as it cascades into revenue for them, regardless if the videos themselves driving the initial traffic are monetizable are not. This allows YouTube to escape regulatory scrutiny by claiming — in the event of a Congress hearing — that videos with problematic content do not accrue them any revenue, and are therefore a financial liability. But what the data seems to suggest, is that even when YouTube pushes a narrative of a squeaky-clean façade to the public, the overall gain to their public brand is very easily offset by the amount of non-corporate, occasionally-ardent content they’re able to consume on the platform. People still watching basketball highlights on ESPN’s channel does not pose a threat to the more clandestine parts of the business.

In the average layman’s terms, that means that YouTube promotes properly-vetted content on its prized Trending tab as a way to obfuscate its darkest corners from advertisers while directly making revenue from it, and in the case of YouTube’s least-prideful creators, the platform uses them as a hook to stay and hopefully dig their claws into what the platform can extract most revenue from. This discrepancy however, has given rise to some of the vilest forms of speech atop the YouTube charts, powered by a well-documented phenomena referred to as the “rabbit-hole effect” — further polarizing viewers — all-the-while those working to stifle its spread are simultaneously getting viewership ripped away from them due to YouTube’s inability to discern endorsement from rebuttal, and through the platform’s glaring preferential treatment of politically conservative content, over its progressive counterbalance.

The rabbit-hole effect is especially interesting to single out because it’s the one element that’s remarkably opaque, but yet, we can see through the outcome of the clearest. After I published my story on YouTube’s erratic decision-making and how it often ends up tipping the scales in favor of conservatives, I got loads of responses from people who all seem to have one consensus in mind–despite their professed disinterest in extremist far-right content, (either through the platform’s own tools or their watching habits) they’ll still get slapped a Ben Shapiro or Steven Crowder video on the recommendation bar more than say, yet another progressive YouTuber. That seeming preference with which YouTube treats extremist content on the right can at least partly explain why even the shabbiest contributors are pulling in way ahead of some of the biggest players on BreadTube.

Anecdotal evidence can be quite tricky to evaluate, but when it culminates into a sweeping trend where even those who opt not to partake into YouTube’s corner of the alt-right still get recommendations for it, it starts to put into real question what YouTube’s real motivations are even as that content continues to produce tangible real-world harm.

The issue with pointing out that YouTube prefers to promote certain political views over others is fairly-obvious–no two creators on the opposite ends of the political spectrum are the same, so trying to draw any statistical link is nigh-on impossible to do. But what’s indisputably evident, is that we haven’t had massive shooting from a self-professed progressive who’s dug in their heels as deeply in their idols’ messaging as say, the Poway Synagogue shooter, or Christchurch Mosque shooter did in far-right ideals. Anti-Semitism, Islamophobia, homophobia, and all norms of bigotry customary to the platform have found a welcoming home in YouTube, and it’s unclear despite the platform’s repeated pledge to do better, whether it’ll actually make much of a difference now that most of the damage has been already done.

It’s difficult to get politicians to agree on what interpretation is most sound for YouTube’s current woes. Some would argue that right-wing bias isn’t an issue at all. The prospect of getting a divided Democratic-majority US House of Representatives — much less a Republican-majority Senate — to agree on a unified set of guidelines to tackle online radicalization is close to null, and if such efforts were ever to get off the ground, the GOP would still likely invoke the spirit of the First Amendment even though it does not apply in the case of a private company enforcing its own rules of content moderation within the borders of what it conditioned its users to agree upon in the first place. Positing that YouTube’s actual problem is a bias against conservatives, or a not-loose-enough system of moderation undercuts the very basic fact that the platform created rules so that they may be enforced in the first place, and when it forces users to remain engaged the platform even though it does not ratify its end of the user agreement, that’s called a platform monopoly. YouTube in that way is forced to either self-regulate, or have the reigns of its arrogance pulled back by a responsible legislature.

Signs of that changing even under the policy-incompetent Trump administration are starting to make way through the bureaucratic hellscape that is the US Department of Justice and the FTC. The DOJ is going after Google in an antitrust probe, claiming its businesses have near-decimated any potential for competition–with an emphasis on their most-successful parts of the business, i.e. Google Search, and YouTube. The FTC on its part is investigating claims that YouTube’s algorithm has become a discovery gold-mine of videos featuring minors, prompting major concerns about the potential problematic use of such a feature — citing specifically the rabbit-hole effect — to the benefit of child predators.

These are issues that the bipartisan makeup of America’s driving force in the legislation space can agree upon, but they’re far and away only but a small band aid applied to a massive wound that’s constantly spilling blood even as all involved parties are well-aware of more immediate solutions to stop the bleeding. What’s sure is there’s no real will from the Trump White House to drive the initiative on solving these issues. The closest we’ve ever gotten is the White House hosting a social media summit that does not include any policy experts, or researchers in the field of social media for that matter. So with no shortage of ambivalence in the highest ranks of government, it may all have to come down to YouTubers’ individual ability to study the algorithm and do the best job of gaming it, otherwise, they’ll get buried under piles and piles of extremist and incendiary content that seems to consistently float atop despite YouTube’s best attempts to make it not so.