TikTok: how UX and AI power the future of information (deep dive notes)

This article is a summary of a 1.5hr deep dive session on TikTok from the AI Academy club, written by club member Helin Yontar. You’ll find explanations on why TikTok has created a huge impact, its challenges, and the results of our brainstorming session.

Every day more content is produced than we’ll ever be able to consume. Social media platforms have to find smart ways to recommend information to users to stay relevant. Like any other social media platform that uses recommender algorithms, TikTok uses AI to recommend the best content for its one billion monthly active users each month. Because of its success, the platform has risen to the top two in the ranking of web traffic in recent months.

What makes TikTok different from others?

TikTok's algorithm seems to be especially good at finding whatever niche content you may enjoy. How can it be so good? The Wall Street Journal has conducted research to find that out by creating automated accounts that watched thousands of videos on TikTok. According to the results, before the algorithm “knows” you it tends to recommend you popular videos with an average view count of 6.3M. After a while, the algorithm started recommending content that was much more specific to the automated accounts’ pre-set interests, with an average view count of 0.78M. On average, the algorithm needed less than 2 hours to “know” the accounts. In some cases, only 40 minutes.

One of the outcomes of the research was that the most important data point for the algorithm was the time spent on each piece of content. This is the goal that the algorithm seeks to optimize, and the entire TikTok’s UX is designed to measure that effectively. The outstanding performance of TikTok’s algorithm is the fruit of the perfect marriage between UX and AI.

At the end of this deep-dive session, we acknowledged that TikTok’s AI and UX redefined how we consumed content and asked each other the following question: How can TikTok change other industries negatively or positively?

We started with commerce. The TikTok Feta effect showed how TikTok can have an influence on consumer behavior. After a feta cheese pasta recipe became viral on TikTok, feta became the first search term on the Instacart grocery delivery app and the demand for Feta cheese was up 200 percent. All this public attention being directed to one thing for a short amount of time creates fake scarcity moments: things that would normally not have scarcity for suddenly became scarce, but just for that brief virality moment.

The Feta cheese TikTok led thousands of people to go to the supermarket and buy Feta cheese. If these people had the chance to buy it on TikTok right after watching it, they would probably do so. This effect paints a promising picture for the e-commerce sector, but there are some caveats. AI Academy club member Francesco pointed out that TikTok’s UX may not transfer well to an e-commerce context as buying behavior patterns for buying products are different from consuming online content. TikTok would have to use your content preferences as a proxy for buying interests, which is not granted to work.?

We spent a lot of time thinking about how TikTok can redefine society. The first worry started from this sub-question: what if the recommendations are too good? Gen Z users watch an average of 233 TikToks a day. These represent a lot of data points for the algorithm, that can understand deeply what content is most likely to keep the user engaged. However, the algorithm unfortunately can’t consider if the content is good or bad for someone. Machine learning algorithms are designed to optimize for a measurable outcome (in technical terms, a label). The label that TikTok’s algorithm considers is whether you watch a piece of content or not, but what happens when someone who is depressed falls down to a rabbit hole of constantly watching depressive content? The algorithm was successful in optimizing its metrics (view time and retention), but at what cost for the user??

Another aspect we considered is the promise of virality. TikTok newsroom said that neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system. In short: anyone can go viral. An example is Khabana Lane, who in a single year reached tens of millions of followers, and brand deals with companies like Hugo Boss. There are some obvious positive effects in giving recognition to hidden talented creators, but we acknowledged that it has a negative side as well. Some club members reported knowing teens who spend a lot of energy and time posting videos hoping to go viral. While that’s technically possible, it’s also statistically unlikely and potentially ephemeral. Constantly failing at getting that recognition can frustrate youngsters and make them question their self-worth.

TikTok influencer Jhabana Lane on the left and top model Kendall Jenner on the right, in an image part of the Hugo Boss rebranding.

The last challenge we explored is that with super-tailored content you may risk having fewer opportunities to discover new areas of interest. Constantly consuming content on whatever the algorithm narrowed down the interests to, decreases the probability of discovering other interests for younger generations. Some governments are taking notice of these mental health risks. China is a frontrunner: they recently decided to block social media apps for users under 18 between 10 pm to 6 am and after 40 minutes of usage.

Want to watch the recording of the deep dive (1.5hrs)? Subscribe to the AI Academy club here. The first 14 days are free, giving you access to all the other deep dives on topics like self-driving cars, the Metaverse, and more.

Special thanks to the participants of this session: Regina Brix, Helin Yontar, Francesco Bellanca, Francesco Marzotto,