How TikTok decides who to make famous

Part 2: The TikTok System

So what do we mean when we say “then, the post is reviewed by an AI?” TikTok is the most extensive content moderation system that has ever existed.

The 411 on content moderation

To make that claim, we need to understand the status quo in moderation today. Content moderation was one of the first product problems of user-generated content that computer-vision scientists were tasked with solving (ie. filtering out porn). With very little content moderation on one end of the spectrum (like 4chan) and heavy moderation on the other end (like TikTok), “what’s the right balance?” is a complex product question that touches all the major platforms today — especially when you as a consumer might not even be aware that it’s being filtered out because of the “content bubbles” we live in.

From a product perspective, designing for content usually only involves three variations on a theme:

  1. Search: goal-oriented, I know what I’m looking for (Google)
  2. Browse: aimless, not sure what I want, anything good? (Netflix)
  3. Contextual: finding something else along the way (Wikipedia)

These systems are present in some form or another in almost every piece of software we use. When browse-based systems are the priority in a product: algorithmic feeds or “discover” screens (like the FYP), the possibility of users living in a content bubble at scale is inherent in the design. We know that if platforms can influence your consumption of content if it’s viable and profitable to them… then, of course, they can influence public opinions.

Due to these content bubbles (like the FYP) In the last few years, the word “algorithm” has worked its way into the vernacular of non-nerds talking about their Facebook feeds, why their Insta post isn’t doing well or what Netflix and Spotify are serving them up to enjoy.

Read part one of this post, “Leveraging TikTok for growth,” on TechCrunch.

Censorship in Entertainment

When it comes to entertainment, content moderation has an interesting history. TikTok, after all, is a social entertainment platform — it’s not a news service that tells you what is happening in the world like Twitter or a visual social network depicting which fire party your friends attended (Instagram). Entertainment is designed so that we can lean into a mood or change our mood. This is why it’s so hard to pick a movie with friends: what mood are you all feeling/do you want to feel?

Not too long ago, comics and movies were heavily moderated in the U.S. The Motion Picture Production Code lasted from the 1920s to the 60s and the Comics Code of the 1950s regulated the themes and storylines young Boomers were allowed to read.

But seriously — if TikTok is the entertainment network of the next generation, we should all be questioning these things. Not just the fact that TikTok is under national security review for its influence on America, but because they are the very questions we as product people, scientists and engineers need to consider when we sit down at our jobs to create our futures with code.

Shadow-banned content on TikTok

First of all, it’s massive. TikTok moderation is happening on a huge scale with “shadow banning” for 500 million users. As the BBC reports,

Shadow banning is the act of partially blocking content so it doesn’t reach the platform’s entire community of users. It will not be obvious to the user that the creator’s content is not being promoted.

So, what’s getting shadow banned?

  1. Political content. While a lighthearted joke or a meme will probably be fine in the U.S. — political content is allowed on the viral video app as long as it’s ‘creative and joyful’ — plenty of things get removed, especially if they don’t please Beijing, the Guardian recently reported. Here’s a BBC World piece on exactly that about the allegedly shadow-banned Indian TikTok star Ajay Barman for making videos on the theme of Hindu-Muslim unity. What to moderate (but do we really mean censor?) in the U.S. is “still in question” and TikTok has hired very nice lawyers to frame this in a positive light.
  2. Pro-LGBTQ+ content. Tiktok censors pro-LGBTQ+ content. On a purely scientific level, identifying content that should be removed from platforms (ie. child abuse and porn) is still a difficult computer vision problem (which we go into in more detail below) and as a result, children doing adult things or being abused often make it through the same AI fence that blocks legitimate LGBTQ+ expression. However, given the list of things being moderated here, it’s a reasonable assumption that human moderation is at work shadow-banning this content too.
  3. People with disabilities and marginalized individuals. Recently, TikTok moderators were told to hide all videos made by marginalized people. TikTok moderators say they are doing this to “prevent bullying,” but the world loses brilliant content and is ultimately a less diverse place when bullies get to be tastemakers.
  4. “Anything they don’t like.” Tiktok also censors at its discretion, for example, this Turkish moderator’s account of “overtly gay content, politics, religious content or drinking would be chief reasons to take a video down” (alcohol and homosexuality are legal in Turkey), not to mention content of villages and shantytowns they didn’t like the look of.

A political/moral content experiment

@davidjustinndark humor for you guys 😈🥵 #itsajoke #chill #foryou #comedy♬ Thats the whole point – damn_it_cas

While testing what makes content pop, we did a little growth-hack experiment using content that touches “political” and “moral” topics just to see what would happen. Recently, this TikTok mocking a pro-life bumper sticker appeared on our growth expert’s FYP.

He got curious about this because we know that TikTok censors political content. So, true to how we just walked you through making a TikTok, he copied the format to make his own.

Rather than use the version of Lil Dicky’s song “Pillow Talk” from the original video, he opted to create a shortened version for his post. While still the widely familiar song, it would receive higher completion rates and loops. The original post referenced a pro-life bumper sticker, but for his experiment, he decided to create a fake, yet believable, Trump tweet with the same message. This not only further politicized the content but crossed the boundaries of fake news and misinformation.

The results: 500k views, 70k likes and over 50 recreations in 3 days. It made it onto the FYP. So, while TikTok is clearly moderating a lot of things, it’s interesting to see what slips through the cracks in this particular experiment involving fake news.

What this kind of moderation means for creativity

While TikTok has weird and unexpected videos, the aperture for what constitutes weird and unexpected still only allows for a narrow set of things that TikTok has trained its AI and human moderators to recognize.

A certain middle of the road flavor of not-rocking-the-boat entertainment is the norm here. While some of it is incredible content and we respect the creators who make it, it’s important to understand the leash on creativity and what is okay here, especially if creative entrepreneurs are the new artists. If so, art as we know it is about to become transform as it potentially ceases us to ask tough questions in ways that are visually, sonically and emotionally arresting.

It’s important that there are entertainment platforms that put creativity first like Byte.

Beyond content: moderating behavior

The social credit system. Part of TikTok’s appeal is that it feels like “the last sunny corner of the internet,” according to The New Yorker. But how do TikTok moderators and AIs keep it that way? Part of the strategy is adhering to social and cultural norms in the local region. An emerging approach is the social credit system pioneered by municipalities in TikTok’s home county.

This system awards greater freedom to participate in society (in this case TikTok’s platform) to people who don’t break the rules or make other peoples’ experiences worse. This clearly influences the content moderation and FYP promotion (as discussed in Part 1, the review process). In China, this system is even starting to cross the boundaries from virtual to real. TikTok has partnered with local courts in Nanning, Guangxi to “display photographs of blacklisted people as advertisements between videos, in some cases offering reward payments for information about these people’s whereabouts that are a percentage of the amount of money the person owes.”

While using social media to guilt people into paying debts seems pretty heavy-handed, spotlighting social behaviors that have consequences in the digital town square also promotes trust among the people who live there. Maybe the Chinese approach to being a decent digital citizen deserves some credit? Moderating behavior with rules vs. complete freedom of expression has been a major existential issue for people working at social platforms.

While we are considering AI-supported content moderation at TRASH for certain things (for example, porn), there are a few computer vision pitfalls we’re trying to avoid:

  • Person re-identification: being able to find the same person in all the videos they appear in. This is something TikTok’s parent company ByteDance has invested considerable R&D in over the last few years and is making it available to Chinese police services. A few days ago, the New York Times reported on a similar service, Clearview AI, that over 600 law enforcement agencies have started using in the last year.
  • Gender, race and “criminality” prediction: Unfortunately, this tech is plagued by training data bias and is very difficult to debug. As a result, people from certain groups, especially women and minorities, are often confused with each other. We don’t want to get into the game of accidentally misidentifying the wrong people. Even classifying whether a person in a video looks like they are committing a crime is as likely to be mistaken as correct at this time.
  • Tracking social credit: Content moderation is incredibly difficult. Making a fun, safe space is as hard online as it is in the real world. While we use some AI to determine video content at TRASH, we aren’t classifying people’s intentions or calculating social scores.

The Future of the video internet

TikTok managed to create a ridiculously fun place on the internet. As video becomes the dominant media and streaming tips over 70 percent of global mobile traffic, the choices of TikTok and other large UGC media platforms will have an immense impact on our culture. While the current political climate of “fake news” might be bringing more awareness to how we discover and consume content, the topic is still under-appreciated. Tarleton Gillespie (Microsoft Research New England, Cornell University) in his book “Custodians of the Internet,” says:

Content moderation receives too little public scrutiny even as it shapes social norms and creates consequences for public discourse, cultural production, and the fabric of society.

In the next 5 years, we’ll feel the effects of AI in our life indirectly through the videos we watch every day. It’s important for all of us to know what future is coming for us and to be thoughtful about how we build for it and consume it, even when it seems like just a meme.

Read part one of this post, “Leveraging TikTok for growth,” on TechCrunch.