The feed that knows you

The first time you open a new social network, the platform knows almost nothing about you. The first few minutes can therefore feel strangely generic: a popular song, a football clip, someone cooking, a celebrity, a joke, perhaps a piece of news. You skip most of it, watch a few things for longer, and maybe like one or two posts, but while you are deciding whether the app is interesting, something else has already begun: the app is deciding what might be interesting to you.

At first, its guesses are crude because there is little information to work with. You watch a video about a dog, but that does not necessarily mean you love dogs; perhaps the dog was doing something funny, perhaps you liked the music, or perhaps you simply forgot to swipe. One action tells the system very little, but dozens of actions begin to form patterns, and hundreds can reveal preferences that you never explicitly stated.

You do not need to complete a questionnaire saying that you enjoy Formula 1, architecture, Japanese food, astronomy, or political satire. Your behaviour can provide clues through what you watch, skip, replay, search for, like, share, save, or comment on, while the people you follow and the accounts you visit provide additional information. The precise signals and algorithms differ between platforms, but the general principle is simple: behaviour leaves traces, and those traces can be used to make predictions about what you may want to see next. This is the first thing to understand about a modern feed: it is not simply a window through which you look at the internet. It is a selection, because there is far more content available than could possibly fit on your screen, and something must decide what deserves the next position. Once that selection becomes personalized, two people can open the same application at the same time and enter remarkably different versions of it.

The process is circular: your behaviour changes the feed, the changed feed influences your next behaviour, and that new behaviour gives the system more information with which to change the feed again. The process becomes clearer with a simple example. You watch a video about a new telescope, so the platform may show you another astronomy video, and if you watch that one too, it may test something more specific: perhaps black holes, astrophotography, space exploration, or the possibility of life on other planets. If you consistently ignore astrophotography but stop for black holes, the system has learned something more precise than “this person likes science,” and after enough interactions your feed may begin to contain a surprisingly narrow stream of content that matches a preference you never consciously described.

This can be extremely useful. Before personalized recommendation, finding a specialist lecture, an obscure musician, a small creator, or an unusual hobby often required knowing that it existed and knowing where to search for it. A good recommendation system can cross that barrier for you, introducing subjects, people, and ideas that would otherwise remain invisible, which is one reason personalized feeds can feel less like a catalogue and more like discovery. The problem is that discovery and capture can use the same machinery. The system does not need to understand why something is good for you in order to learn that you respond to it, because your behaviour can reveal that a subject holds your attention without revealing whether you later considered the time worthwhile. A recommendation can therefore be extraordinarily accurate in one sense—predicting what you will watch—while being completely ignorant in another—knowing whether you will be glad you watched it.

This passage is from Scroll on Purpose. Read the introduction and first chapter, browse the guides, or see buying options.