Skip to main content
From Netflix to TikTok: How AI Knows What You Like
On this page

It feels like magic, but it isn't

You open Netflix and the first row is exactly what you were in the mood for. Ten minutes on TikTok and the videos match your sense of humour a little too well. It's tempting to think the app knows you. What it actually has is a very large record of what people like you did, and a system that is good at finding patterns in it.

These recommendation systems decide what you see on YouTube, Spotify, Amazon, Instagram and most news apps. Here is how they work, in plain language.

What a recommendation system does

A recommendation system predicts how likely you are to enjoy, click on or buy something you haven't seen yet, then shows you the items with the highest predictions. Nearly all of them are built on two ideas: collaborative filtering and content-based filtering.

Collaborative filtering: people like you liked this

Collaborative filtering is word of mouth on a massive scale. Say you and your friend Alex both love science fiction, and Alex has also watched a fantasy series you haven't seen. If Alex loved it, there's a decent chance you will too.

The system does the same thing with millions of people. It compares your history (what you watched, liked, bought or skipped) with everyone else's, finds the people whose taste overlaps with yours, and recommends what they enjoyed that you haven't tried yet. It never needs to know what an item is about, only who liked it.

The approach became famous through the Netflix Prize, a competition Netflix ran from 2006 to 2009 with a one-million-dollar reward for anyone who could improve its rating predictions by 10 percent. The winning teams relied heavily on matrix factorisation, a technique that boils millions of ratings down to a small number of hidden "taste dimensions" for every user and every film.

Content-based filtering: more of what you liked

Content-based filtering looks at the item itself. If you watch a lot of crime dramas with a strong female lead, the system looks at the features of those shows (genre, cast, themes, pace) and finds others that share them. It's the logic of "you liked this thriller for its twist ending, so here are ten more thrillers with twist endings".

Its strength is that it works for brand-new items nobody has rated yet. Its weakness is that it rarely surprises you.

In practice: both, and a lot more

Large platforms combine the two and add many more signals. TikTok, for example, looks at how long you watch a video, whether you watch it again, whether you share it, and at the video's caption, sound and hashtags, then blends all of that into a single prediction for each candidate video.

How the system measures "similar"

Both methods depend on measuring how alike two things are. Two users are similar if they rated many of the same films the same way. Two videos are similar if their hashtags, sounds or visual style overlap, or if the same people tend to watch both. How many people watched both Stranger Things and The Umbrella Academy? If that number is unusually high, the two shows sit close together on the system's map, and liking one makes a recommendation for the other more likely.

Repeat that for every pair of users and items and you get an enormous web of associations. That web is what makes recommendations feel personal.

Why it works so well, and sometimes too well

The more you use an app, the more data it has and the better its predictions get. But there's a catch. A system that only shows you what you already like can trap you in a filter bubble, where everything in your feed confirms what you already think and enjoy. That's why a feed can feel perfectly tailored and oddly repetitive at the same time.

Platforms know this, and most deliberately mix in some exploration: items the model is less sure about, to test whether your taste is broader than it assumes. New users face the opposite problem, known as cold start. With no history to go on, the app falls back on whatever is popular until it learns more about you.

You're training it all the time

Every action feeds the system: a series you finish, a video you skip after two seconds, a song you replay. You are constantly teaching it your preferences, mostly without noticing. That raises fair questions about privacy (what is collected and how long it is kept), bias (whose taste the system learns to favour) and autonomy (how much of what you choose is really your own choice). Understanding the basics is the first step to having an opinion on any of them.

Taking back some control

  • Clear or pause your history when your recommendations have drifted somewhere you don't want them.
  • Seek out a wider range of content on purpose. The algorithm can only show you what it thinks you might like.
  • Use separate profiles for different people, or even different moods, so one doesn't pollute the other.
  • Use the "not interested" button. Negative feedback is one of the strongest signals most systems get.

Can you switch personalisation off altogether? Some platforms let you limit tracking or offer a feed that isn't personalised, and in the EU the Digital Services Act requires the largest platforms to offer at least one recommendation option that isn't based on profiling. Expect less relevant suggestions if you use it.

Final words

You don't need to be a data scientist to understand recommendations, just a bit of curiosity and a healthy dose of skepticism. These systems are very good at predicting what will keep you watching, which isn't always the same as what is good for you. Use them with that in mind, and they work for you rather than the other way round.

Share

Related articles

Leave a Reply

Your comment has been submitted. Thank you!

Comments