Why Personalized Feeds Can Keep Pulling You In

A good bookseller will put a novel in your hands and say, “If you liked the last one, you’ll like this.” They are usually right, which is why you trust them. Then, at six o’clock, they turn the sign on the door, and the recommendations stop, because the bookseller has a train to catch and the shop has a closing time. Nobody has ever been recommended a book at half past two in the morning by a bookseller. It would be alarming.

You will open the app tonight for a minute, between the end of the dishes and the start of the evening. The first clip will be exactly your sort of thing. So will the second. You will think, faintly flattered, that it really does know you, and you will be right about the dog videos and wrong about everything else. The feed will not close at six. It will not close at all. When you finally look up, the radiator will have gone cold and the lamp across the street will be the only one still on, and you will not be able to name a single thing you watched. Or you will put it down after the first one. That part was never written by them.

Why “it knows me” feels like the whole story

The tailored feed is the most flattering design in the building. Each clip arrives pre-chosen from what you watched, paused on and skipped, so the screen seems to be holding up a mirror, and most of us find mirrors hard to walk past. Reviewers of platform design list personalized recommendations alongside endless feeds as features built to prolong engagement, which is a dry way of saying somebody tested them against your bedtime and they won [3].

The YouTube watch-next chain works on a choice you made: you searched for a guitar lesson, and the sidebar took it from there. A short-video feed skips even that courtesy. There is no first video you chose. There is a screen, already playing, already guessing, and a thumb. In a survey of 120 YouTube users, autoplay and recommendations were the features that most undermined people’s sense of being in charge, while search and playlists supported it [2]. The short-video feed is, more or less, recommendations and autoplay with the search bar moved to a back room.

So the obvious story goes: the algorithm knows you too well, and that is why you cannot stop. It is a good story. It casts you as a person so interesting that a supercomputer had to be built to keep up. A study published this month suggests it may be the wrong way round.

What the eye-tracking study actually found

Researchers in China and the United States compared people with higher and lower scores on a measure of problematic short-video use [1]. Each participant supplied screenshots from their own recommendation feed. Then came the clever part: each person’s personalized screenshots were used as somebody else’s non-personalized ones, so the two kinds of image were the same pictures, just shown to different people. That removes the obvious objection that one set of clips was simply more eye-catching.

Same screenshots, different eyesHeavier users looked at anyone's clips almost as readily as their own

Lower problematic use

Attention picks favorites

• Bias toward clips from their own feed
• Gaze went to what was theirs
• The tailoring did the pulling

Higher problematic use

Attention stops being choosy

• About the same bias for their clips and a stranger's
• Less selective gaze on the eye tracker
• The format did the pulling

Participants did a dot-probe task, a standard way of measuring where attention goes first: two images flash up side by side, then a dot appears in the place of one of them, and quicker responses to the dot reveal which image had already caught the eye. An eye tracker recorded where people actually looked.

The result was an interaction, which is research for “it depends who you are.” People with lower problematic use showed a selective bias toward their own personalized content: their attention went to the clips that were theirs. People with higher problematic use showed about the same bias toward personalized and non-personalized clips alike. Their gaze was less selective. Across the sample, higher problematic-use scores went with lower selectivity [1].

Read that slowly, because it turns the flattering story inside out. For the lighter users, the pull was about the tailoring. For the heavier users, the tailoring seemed to matter less, because short-video content in general had become the thing that caught the eye. The authors describe the bias as “generalizing beyond personally tailored cues.” In plain terms: the feed does not need to know them very well anymore. Anyone’s feed will do.

That is a less romantic finding than “the algorithm knows my soul,” and a more useful one. A feed that only works on you when it guesses right can be outwitted by making it guess wrong. A pull that has spread to the whole format will survive a blander feed.

Where the evidence runs out

Where the evidence runs out: this is one study, published online ahead of print, and the abstract does not say how many people took part, which is the first thing I would want to know. People were grouped by habits they already had, so the study cannot say which came first: whether heavy use flattened their attention, or whether people whose attention was already less choosy drifted into heavy use. Participants looked at screenshots in a lab, not at a live feed in bed, and a dot-probe task measures where attention goes in the first fraction of a second, not why someone keeps swiping for forty minutes. The authors themselves call attentional selectivity a “potential” process linking personalization and problematic use. It is a mechanism worth watching. It is not proof that algorithms cause addiction, and anyone who tells you otherwise has read the headline and not the conclusion.

What it does do is retire one comforting idea: that the problem is the cleverness of the recommendations, and that a heavy scroller is simply a victim of very good guessing. The guessing matters. So does the format: the swipe, the autoplay, the absence of a last page. Those were decisions too, and they were made in rooms where nobody was asked what time you get up.

What the scroll is standing in front of

Here is the less comfortable part. The gap between the dishes and the evening is a seam in the day, and seams are where the feed lives. What would otherwise fill it is often small and slightly dreaded: the reply you owe, the form you have not opened, the bath that means the day is over, the conversation in the next room that would take more than a thumb. The tailored feed is a very pleasant place to wait for the feeling to pass. The feeling, as a rule, waits longer.

Some nights that is fine. Some nights the dog videos are the kindest thing in the room, and nobody needs to apologize for twenty minutes of a golden retriever failing to catch a ball. The question is only whether you chose the twenty minutes or held the door while they were taken.

A fictional example

The following is an illustrative example, not a real person’s story.

Megan opens a short-video app at 10:40 p.m. on a Thursday, meaning to watch one clip while the kettle boils. She has read that the algorithm “knows her,” and on a bad week she finds this almost touching. In the spirit of an experiment, she spends a Saturday afternoon retraining it: she taps “not interested” on anything in her usual groove and lets the feed fill with gardening, which she has no feelings about at all.

On Sunday night the feed is duller, and she watches gardening for thirty-five minutes. She does not know the name of a single plant afterward. The tea is cold. The email she meant to answer is still unanswered, and it was always the email.

What the example shows is the study’s twist in miniature. Making the feed guess wrong did not loosen the habit much, because the pull had moved from the particular clips to the shape of the thing: the swipe, the next one, the absence of an end. A blander feed is still a feed.

Putting an ending back

1. Decide your exit before the first clip. A number of clips or a clock time, chosen with the app still closed. This helps with the format’s main trick, which is that the next one is always already there.
2. Treat “it knows me” as a sales line. When a clip feels uncannily yours, say so out loud, then ask whether you would have searched for it. If the answer is no, the feed chose it, and you are allowed to decline.
3. Retrain the feed if you like, but do not rely on it. A reset is a cheap, reversible experiment. If the pull is in the format rather than the tailoring, it will not do the stopping for you.
4. Go in through the search bar. Search and playlists supported people’s sense of agency in the YouTube research [2]. Looking up the one thing you wanted gives the session a shape that a feed never will.
5. Name what the seam was for. Before you open the app between two parts of the evening, say what comes next. If it is the email, open the email first and give the feed whatever time is left after it.

Try this for a week

Each night, before you open a short-video app, write down your exit: a count or a time. Afterward, note whether you kept it and roughly what you watched. At the end of the week, compare the notes with your screen-time report. If the nights you kept your exit were the nights you remembered what you watched, that is your answer about whether you need the tailoring at all. If you have tried this already, a stopping point before the video app looks at setting the exit in more detail, and short videos vs long-form content explains why the sixty-second format is so good at hiding its seams.

When to talk to someone. If short-video use is costing you sleep for weeks at a time, if you feel unable to cut back despite wanting to, if it is getting in the way of work, study or relationships, or if the scrolling is covering low mood, anxiety or loneliness that will not lift, talk to a doctor or therapist. Those signs are worth a professional conversation, and asking for one is a sensible step, not a failure.