You never told the app what you like. It watched. When you paused on a video for eight seconds instead of two, that was a signal. When you replayed one, that was a stronger one. When you flicked past in half a second, that was a signal too. It logs all of it.
Flick past1 how loud
Like3 how loud
Watch to the end4 how loud
Replay6 how loud
Share10 how loud
Made-up loudness, real order.
Some signals are louder than others. A share is louder than a like. A replay is louder than a watch. A comment is louder still. Turning the sound on, tapping the profile, taking a screenshot: all counted. A like is one of the quietest signals there is, and the only one you know you sent.
You watch a video twice
vs
You tap like on a video
Which of these does the app learn more from?
Tapping like. A like is you telling the app directly.
Not yet. It is direct, and it is quiet. People like things for friends, for politeness, for a joke. Replaying something is harder to fake.
Watching it twice. A replay is a strong signal, and you did it without thinking.
Right. A replay costs you time and happens on its own. The app trusts what costs you something more than a tap you might give out of habit.
Neither. It only learns from what you search for.
Not yet. Searching is a small part of a session. Most of what the app learns arrives from what you do while scrolling, and you do that far more.
0 in every 100 videos31 in every 100 videos62 in every 100 videos
Session 1Session 5
Made-up: a topic's share of a feed after the reader finishes it each session.
The guesses update fast. Watch three cooking videos to the end tonight, and by tomorrow the score for cooking videos has risen. Nobody decided that. The guess 'this person finishes cooking videos' now has three yeses behind it. The app tries a few more to check. If you finish those, more come.
Move the control to set how long you pause on cat videos, on average. You never tap like on any of them. The rule is made up: three cats in a hundred for every second you pause.
1 of 20
Cat videos in the next twenty.
1 secondPausing 1 second on average: by the next session about 3 in 100 videos are cats, or 1 of the next twenty.
2 of 20
Cat videos in the next twenty.
3 secondsPausing 3 seconds on average: by the next session about 9 in 100 videos are cats, or 2 of the next twenty.
4 of 20
Cat videos in the next twenty.
6 secondsPausing 6 seconds on average: by the next session about 18 in 100 videos are cats, or 4 of the next twenty.
6 of 20
Cat videos in the next twenty.
10 secondsPausing 10 seconds on average: by the next session about 30 in 100 videos are cats, or 6 of the next twenty.
9 of 20
Cat videos in the next twenty.
15 secondsPausing 15 seconds on average: by the next session about 45 in 100 videos are cats, or 9 of the next twenty.
You never liked a single cat video. Why is the feed filling with them?
Because you paused. Pausing is a signal, and you sent it many times.
Right. A like is one quiet signal. Fifteen seconds of pausing, many times, is a loud one. The app follows the loud ones.
Because cats are popular with everyone.
Not yet. Popularity gets a video onto the list. Its place near the top of your list came from your own pauses, which is why the control moves the count.
It is a mistake, and it will correct itself.
Not yet. It is working as designed. The guess was 'this person pauses on cats', and that guess was right. It changes when the pausing does.
Move the control.
A fake accountpauses on sad videos
↓→
Within about an hour
↓→
Its feedmostly sad videos
In 2021 journalists set up fake accounts on a video app. Each account was given one interest, and showed it only by pausing on certain videos. Within about 40 minutes to two hours, each feed was mostly that interest. One account paused on sad videos. Its feed became sad videos.
How many separate signals about you does a big app's system use when it scores one post for you?
Pick one. Nothing is scored.
What you could list about yourselfa few dozen
vsvs
What the app logsthousands
Thousands.Guesses run low because you can only think of a few dozen things about yourself. The app does not think in things about you. It counts every pause, the hour you open it, every video you skipped, and Facebook has said its system uses thousands of these for each post.
Put in order how one evening changes the next day's feed.
Tap them in order, first to last.
Three finished→Logged→Guess goes up→Score goes up→More cooking
Right. The step to notice is the third one. The app does not decide you like cooking. A number goes up, and everything after that number follows it.
A shopperno children
→
The app suggestsbaby clothes
A shopping app suggests baby clothes to someone who has never had a child. What is the most likely explanation?
The app knows something about them they do not.
Not yet. It knows what they did on it. A guess from a pattern is not knowledge; it is a bet, and this one is probably wrong.
They did something that people who buy baby clothes also do, like lingering on a pram, and the guess followed.
Right. The guess is not 'has a baby'. It is 'acts like people who bought baby clothes'. One long look at a friend's pram link is enough to nudge that number.
It is random. Apps show everyone baby clothes sometimes.
Not yet. Nothing on a scored feed is random. A suggestion is a guess with a number behind it, and something moved that number.
Lesson complete
Every pause, replay and skip is a signal, and the guesses about you update from them within hours.