How TikTok Uses AI to Engineer User Addiction
TikTok has over 1 billion monthly active users, rivaling competitors such as Instagram and YouTube. TikTok's artificial intelligence algorithm is key to its breakout success.
If you have just a few minutes to spare, here’s what you need to know:
TikTok uses a best-in-class AI-based recommendation algorithm to drive user engagement. This algorithm is better than that of its competitors such as YouTube and Instagram.
TikTok does not require the user to declare explicit interests. Users have the option to declare interests, but, even if they do not, users will find that the app quickly starts recommending the right videos within a short time after they start using it. This makes the TikTok experience seem magical.
TikTok’s AI tracks each and every behavior as users watch the recommended videos. Then, with the information the app tracks, it maps out the user's personality — highlighting emotions such as sadness, happiness, and apprehension and other interests and hobbies.
TikTok also uses AI to enhance the user experience of both consumers and creators.
The addiction TikTok engenders in users is a testament to the progress in AI but also raises concerns about user privacy and safety.
This post is sponsored by Multimodal, a NYC-based development shop that focuses on building custom natural language processing solutions for product teams using large language models (LLMs).
With Multimodal, you will reduce your time-to-market for introducing NLP in your product. Projects take as little as 3 months from start to finish and cost less than 50% of newly formed NLP teams, without any of the hassle. Contact them to learn more.
Ankur’s Newsletter is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
Ever heard of 15 minutes of fame? It's out of vogue now — 15 seconds of fame is the new norm. Thanks to TikTok, a Chinese video-sharing social media app, short videos featuring music, dance, information, and acting are one of the fastest-growing forms of entertainment today.
TikTok's viral success is one-of-a-kind for a Chinese app. That's not to say that other Chinese apps have not risen to popularity before. Apps like Alipay, Tencent, and Baidu are well-known, too. But their popularity is limited to the Chinese mainland. TikTok, however, attracts a truly global audience and has a substantial base of users in the United States.
While there are several reasons behind the app's meteoric rise to popularity, its fascinating — and in parts dangerous — algorithm is the key. Let's explore TikTok, the AI that supports it, and how it impacts users today.
How TikTok Rose to Fame
Listen to “Old Town Road” by Lil Nas X. Or “Oh No” by Kreepa. Or “Say So” by Doja Cat. These are all fantastic, catchy tunes. Do you know what else they have in common? These, and hundreds of others like these, all rose to popularity on TikTok.
TikTok is a short-form video-sharing app owned by the Chinese tech company ByteDance. The app enables users to create 15-second videos with built-in music, filters, and background. As a creator, you can make videos in various genres, including humor, prank, gaming, art, healthcare, and more.
Within China, TikTok is known as DouYin. And even though the parent company and user interface of DouYin and TikTok are the same, their servers are different. If you use TikTok, you cannot access content from DouYin and vice versa.
Although TikTok was released globally in 2018, its history dates back to 2014. Alex Zhu and Luyu Yang were two friends working on an EdTech app to make it easier for students to learn and study at home. This app would also give talented teachers without a traditional pedigree an opportunity to become popular.
However, things did not go exactly as planned. Even while being developed, the app’s purpose started tilting heavily towards entertainment. It was soon released under the name Musical.ly and became wildly popular in a short span of time. By 2016, it had over 70 million users creating and sharing music, lip-sync, and humor-based videos.
ByteDance was developing DouYin around the same time. DouYin, too, became very popular, registering millions of downloads within a year. Seeing Musical.ly's popularity, DouYin acquired Musical.ly and merged it with itself in 2017. A year later, TikTok, the combined app, was released globally.
Since 2018, the app has seen steady growth in global markets like North America, Southeast Asia, and Western Europe. Today, it has over 1 billion monthly active users (MAUs) and realizes year-on-year growth of 45%.
To put this in perspective, the American social media app Instagram has around 1 billion MAUs. Considering how new TikTok is to the scene, the fact that it competes so well with giants like Instagram is impressive.
But what gives TikTok such an edge over its better entrenched and financially stronger competitors? Yes, TikTok offers built-in music and filters that are easy to customize compared to apps where users have to do all the work from scratch. But Instagram and YouTube have also started offering similar short-form content creation tools. On this measure, TikTok is not wildly different from its peers.
What explains TikTok’s meteoric rise? Its algorithm.
The TikTok Algorithm Drives User Addiction
TikTok users spend more time on TikTok than Instagram users spend on Instagram and Snapchat users spend on Snapchat. This might not necessarily be because TikTok’s quality of content is better than that of its peers. TikTok may just be better at presenting the right content to the right user.
The "For You" section of the TikTok app is astonishingly good at recommending videos to users based on their interests. Once users start watching videos on TikTok, users will keep seeing more videos to keep them highly engaged. The better TikTok does at engaging the users with its video recommendations, the more time they will spend on TikTok, eschewing other apps such as Instagram and YouTube.
An Overview of How the TikTok AI Works
Here is how the TikTok algorithm works:
TikTok analyzes user behavior (i.e., interaction with content) and their basic data like language and device used.
Based on this analysis, TikTok forms a detailed user profile, highlighting the person’s core personality traits and interests.
TikTok then uses this profile and matches it with the most relevant content. To find which content is relevant, TikTok uses three things: natural language processing to make sense of the caption, computer vision to identify what the video contains, and metadata such as hashtags, audio used, and people tagged.
When a video is first posted, TikTok shows it to a small audience. Then, based on whether or not the group likes, watches, or shares the video, TikTok decides whether it should recommend the video to more people.
The TikTok AI uses a psychographic predictive model. TikTok uses people's reactions to the on-app content to draw a user's personality. These seemingly scattered and unrelated reactions go a long way in documenting someone's unstated desires, feelings, and thoughts.
To simplify this, think about your current or former partner. While you might have started out as strangers or friends interacting on a few dates and phone calls, that person can now judge and predict how you would react to certain unprecedented situations and people, what you'd feel when something unusual happens, what your moral inclination is, what excites you the most, and much more.
This happens because your partner has learned about you through interactions in the past, even if you did not explicitly tell them every feeling and emotion. Based on what they have learned, they understand your personality.
In the same way, AI learns, too. According to Stanford professor Michal Kosinski, AI can better predict personality traits than humans. It can correctly gauge extraversion, openness, neuroticism, agreeableness, and conscientiousness from unrelated cues.
The TikTok AI gauges these traits by tracking whether or not users "like" a video. By "like," we do not mean the explicit “like” button on TikTok. Even if users partially watch the video or share the video or its creator, TikTok considers it "liked." If the user watches the video in its entirety or watches it several times, TikTok considers the “like” rating to be even stronger. TikTok stores this data and further tests users by showing them more videos of a similar type.
TikTok also studies the kind of device users are using, the language of their device, and their geographic location. These are what TikTok calls "weak signals" contributing to the AI recommendations.
Over a few minutes, TikTok accumulates a large number of user reactions. Then, TikTok forms the user profile, potentially containing detailed predictions about the user’s hobbies, likes, emotional triggers, habits, etc. This information is enough for the app to know what users will like to watch and what they will immediately reject.
On the creator's end, TikTok uses the AI to study facial expressions, metadata such as captions and hashtags, and audio and music used to create content. It stores the information it gathers in a repository which it later uses to quickly customize their "For You" feed.
The main goal of the TikTok algorithm, as revealed by an official company document uncovered by New York Times, is "user value," "long-term user value," "creator value," and "platform value."
How it achieves these goals is a well-guarded secret, but the company has officially said that it uses information about likes, shares, hashtags, and the kind of content users watch to drive its recommendations. But a Wall Street Journal investigation revealed much more.
The Wall Street Journal Investigation of TikTok's Algorithm
TikTok has a potent built-in AI that uses a variety of systems like Computer Vision, Natural Language Processing, and typical metadata to detect what kind of content a video has. Then, based on the user's profile, location, and declared interests, the app recommends videos to them.
Even if a user has not specified interests and has not created an account on TikTok, the app's AI quickly identifies their interests — sometimes in less than an hour or two.
TikTok’s recommendations start as random popular videos and get more nuanced as users spend more time. Ultimately, the apps starts showing niche-specific videos and drives the user to watch and re-watch.
Some of the explicit content on TikTok is not moderated, but it gets recommended to users who have similar interests.
The Wall Street Journal created hundreds of bots to analyze how TikTok works and what drives the frenzy behind it. These bots each had a unique location and IP address but no gender. They had specific interests but didn't enter those into the app.
Contrary to what ByteDance says, the investigators found that the app tracks watch time more than anything else.
The TikTok experience starts similarly for everyone. The bots were recommended a series of random yet popular videos and some not-so-popular ones. Then, the AI tracked how much time the bots spent watching a video, whether or not they paused on one or rewatched it, and whether they skipped the video before it had finished.
According to the WSJ:
"Every second you hesitate or rewatch, the app is tracking you." (Source)
The random stream of videos becomes more nuanced as the user progresses. The user sees videos that more closely fit their interests, even if they have fewer views and likes. As the user spends more and more time on such videos, the recommendations get even narrower, showing users very niche-specific videos.
The WSJ calls this falling into "a rabbit hole of very niche content" — an allusion to how Alice fell through the rabbit hole in Alice in Wonderland.
TikTok’s AI is incredibly magical, learning about users’ deepest desires and interests super fast. But, there are downsides to the technology, too. For example, for one of the WSJ bots whose interests were depression and sadness, TikTok took about 40 minutes to figure out what the user liked.
The app showed the bot random popular videos at first. Then, after a few minutes, the app increasingly started showing the bot sad content, some of which were on extreme topics like anxiety and suicide. By the end of the 40 minutes, 93% of all videos recommended to the bot were depressing. Most of the remaining 7% were just ads.
The WSJ saw similar results with bots with more diverse interests. While it is true that the bots could never be as emotionally nuanced as humans, the results were a good indicator of what the app does to actual users and how it gauges their interests.
Ultimately, TikTok aims to increase user engagement and watch time, even if the app has to recommend non-moderated and potentially harmful content. The danger lies in the fact that some of this content may be pornographic, violent, suicidal, racist, endorsing fake news or propaganda, etc. The problems are compounded by the fact that TikTok has a very young user demographic, which may be more impressionable than the average adult.
However, it should be noted that ByteDance is working on better moderation by using a combination of human screening and, yes, more AI.
Of course, watch time isn't the only parameter TikTok uses to judge its users. The official document uncovered by NYTimes says likes, comments, and playtime dictate overall recommendations.
How TikTok's AI is Different From its Competitors
Unlike YouTube and Facebook, TikTok never asks for user intent explicitly. TikTok doesn’t ask users to click on recommended videos and instead shows them videos immediately in the “For You” feed.
In TikTok, past views, authority, number of followers, and pre-existing popularity of creators don’t play a significant role in a video being recommended.
TikTok nudges creators to become more active by promoting their videos when they become inactive.
Powerful algorithms are nothing new. Apps such as Facebook, Instagram, and YouTube cracked the code of hooking a user long ago. But TikTok goes far beyond just attracting an audience.
According to Guillaume Chaslot of AlgoTransparency, the key aim behind the recommendations is to get users addicted, not to give them what they want.
As Connie Chan puts it,
"TikTok is the first mainstream consumer app where artificial intelligence IS the product. It's representative of a broader shift." (Source)
TikTok is highly engaging because it removes friction from the user experience — users do not need to explicitly declare their interests or vote on what they like in order to see more of what they really want to see. On YouTube, Netflix, or Facebook, you have to search for things, make friends, or interact with posts you are looking for.
On TikTok, users are given a range of tasty dishes on a silver platter. They can eat whatever they like, and they will be served even more of what they like over time.
On Tiktok, creators don't need a specific number of followers or views to go viral. This is the opposite of what is necessary on YouTube. Moreover, users do not even need an account to start browsing through content. It doesn't matter who a user is as long as they either create or watch or do both.
How does a video go viral on TikTok? When a video is first posted on the platform, it appears in the “For You” section of a few users whose past behaviors indicate they might like it. If this guinea pig group doesn't like the video, it doesn't go viral.
Unlike YouTube or Instagram, creators do not need to keep creating content regularly to become famous on TikTok. Some creators have found that if they become inactive after creating content for a while, TikTok automatically promotes their content to call them back to the platform.
As TikTok says:
"While a video is likely to receive more views if posted by an account that has more followers, by virtue of that account having built up a larger follower base, neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system…" (Source)
Moreover, the amount of authority the creator has over a piece of information does not matter for the platform. Creators do not need to be a celebrity or hail from any particular organization to become popular.
The Hidden Dangers of Addictive AI
We are not going to present the algorithms' most apparent risks here. Obviously, prolonged usage of such highly addictive social media can affect vision, mental health, relationships, level of daily physical activity, and overall well-being. Instead, let’s explore the risk of being unfairly influenced through such platforms.
The AI profiling of a user that TikTok does can help it understand why someone invests in a product, how they might vote, what makes them tick, what personal, religious, or professional biases they have, what inspires them, etc. All of this information is sensitive and makes people vulnerable to subliminal advertising and propaganda.
It’s worth mentioning that companies like Facebook and Google also have similar data, and the possibilities of this data being used manipulatively exist with them as well. But since TikTok is better at profiling people, the threat is potentially even more potent.
Such information about high-profile individuals like politicians and government workers can be used to unfairly influence elections. Even in the case of children, the user data can be used to influence their mental and emotional development and moral proclivities or to blackmail them.
The app has already been accused of selling sensitive user data to governmental and nongovernmental agents. These accusations have triggered investigations and bans for the app. But TikTok’s popularity continues to grow, mostly unfazed, a testament to just how addictive the app is.
TikTok's meteoric rise to popularity can be attributed to its incredible algorithm, its entertainment format, and its elimination of friction in the user experience.
The TikTok AI engineers user addiction. The AI does not focus on providing informational value to a user. Instead, it promotes content users watch, even if they dislike it. TikTok does not ask a user to express their choice or intent explicitly. Rather, the AI starts predicting everything with barely any information about the user, making the user experience frictionless.
Some of the non-moderated content on TikTok can trigger vulnerable users and cause mental, emotional, and physical health issues. The longer-term impact of this is yet to be determined.
One of the primary dangers of such AI, whether used in TikTok or other social apps, is that it can reveal deep personality insights, which malicious agents can use for propaganda and subliminal messaging.
Despite having a human-backed filter for violent, hateful, pornographic, and other disturbing content, TikTok fails to prevent such videos from seeping through users’ “For You” feeds. It needs to do a better job moderating such content on the platform, particularly for younger users.
TikTok is undoubtedly a significant technological development in AI and machine learning and serves as a wake-up call for its competitors. Hopefully, ByteDance will proceed cautiously as it continues to develop the AI, focusing more on keeping the platform safe now that it has achieved a truly global scale.
Subscribe to get full access to the newsletter and website. Never miss an update on major trends in AI and startups.
Here is a bit more about my experience in this space and the two books I’ve written on unsupervised learning and natural language processing.
You can also follow me on Twitter.