TikTok’s recommendation algorithm has one job, from the viewer’s perspective: figure out what you want to watch next before you know you want to watch it. You open the app. Within minutes, something you would never have described as an interest is on the screen, and somehow you are watching a man restore a nineteenth-century door hinge.
The For You feed can feel like machine intuition. But a useful recommendation is not evidence that an app understands your personality, reads your mind, or knows what is good for you. Those are rather larger claims than “successfully located another hinge.”
What the Algorithm Actually Watches
In its June 2020 explanation, TikTok described a combination of interactions, video information, and device settings. Likes, shares, follows, comments, captions, sounds, and hashtags all featured. Finishing a longer video was an example of a stronger interest signal than sharing the creator’s country.
The company also said follower count and an account’s previous hits were not direct recommendation factors. That is different from saying followers are useless: an established audience can still supply viewers. Nor does it mean one excellent completion rate guarantees a viral video.
What this explanation did not provide was a universal scorecard saying that a replay is worth five likes or a share is worth three comments. The exact weighting is not a public recipe. Anyone selling certainty about it should at least throw in a complimentary crystal ball.
The Cold Start Problem
A new account gives a recommendation system relatively little to work with. TikTok’s published explanation described starting with selected interests, or popular videos if none were selected, and adapting to subsequent interaction. It also described deliberately mixing in different content to avoid endless repetition.
Researchers can examine this process without having access to every line of the platform’s code. In a 2022 audit, Maximilian Boeker and Aleksandra Urman used controlled accounts to test language, location, following, liking, and viewing behavior. All affected recommendations in their experiments. Following had the strongest measured influence, followed by liking and video view rate.
That is a finding from a particular experiment, not a permanent ranking of every signal for every user. It also illustrates why “likes barely matter” is too confident. A study can show personalization happening without supplying a thirty-minute countdown that applies to everyone.
The Rabbit Hole Problem
The concern is easy to understand. Watching something does not always mean wanting more of it. People watch out of curiosity, irritation, disbelief, or because they have temporarily misplaced the ability to put the phone down. Engagement alone cannot explain which of those experiences occurred.
A feed that keeps serving a narrow theme may be convenient when that theme is sourdough. It becomes more troubling when the material is distressing or misleading. The researchers discussed filter bubbles and problematic content, but their experiment does not prove that every user follows an inevitable path toward more extreme videos.
This is also why questions about musical recommendation and repetition feel familiar here. Personalization can help us find something we love. It can also make the next discovery suspiciously similar to the previous one.
A Prediction Is Not a Psychological Portrait
There is a meaningful privacy question in what a service can infer from behavior. But we should not turn that concern into the unsupported claim that TikTok measures emotional states with known precision, or collects behavioral data faster than every other consumer platform.
A system may predict that you will watch another video without knowing why. Keeping those two things separate makes it easier to ask useful questions about data collection, access, retention, and independent scrutiny. “It knows everything” is an effective horror premise and a poor technical specification.
The Creator Side
For creators, recommendations offer a route to people who have never followed them. That opportunity can also feel unstable: one successful post does not purchase a permanent audience for everything that follows. The algorithm is not a loyalty scheme, even if creators sometimes wish it came with a stamp card.
There is no need to treat the For You feed as either magic or a villain with a spreadsheet. It is a powerful distribution system, built around imperfect signals and business choices. Understanding those limits is more useful than memorizing a supposedly secret formula, and might even help you escape the door-hinge videos before dinner.
