YouTube 的 AI 推荐是怎么跑起来的？What Makes YouTube's AI Recommendations Tick?

The plan was one five-minute clip. Two hours on, you are seventeen videos into a subject that was unknown to you yesterday. None of that is chance — it is an AI system performing precisely the job it was designed for. Below is how the machinery works, explained in plain English.

A quick clip on repairing a dripping faucet was all you wanted. Next thing you know, the evening has vanished and you are deep in Soviet submarine history, with no clear memory of how the descent began. Familiar? Then you have already encountered the internet's most powerful recommendation engine — you simply never caught it in the act.

So what is actually going on? No wizardry is involved, and nobody sits in a room choosing clips on your behalf. What exists is a stack of machine learning models: they observe your behaviour, learn from the millions of other people behaving the same way, and keep recalculating the odds that one more video will hold your attention. This piece pries open that black box and traces how a video earns its place in your Home feed, the Up Next queue, or the Shorts row.

One distinction deserves drawing before we continue, since two separate notions get blurred together constantly. Not every automated decision inside YouTube qualifies as "AI" in the deep-learning sense; a fair amount is plain rule-based plumbing, such as hiding content in a language you have never once watched. Our breakdown of where AI ends and automation begins marks that boundary precisely. The recommendation engine, by contrast, is the real thing: machine learning trained on a volume of viewing data — billions of watch sessions — that is hard to picture.

01The Plain Version: Behaviour Teaches It, Not Declarations

Think of the system as deaf to your preferences and blind to your intentions; only conduct gets recorded. Every tap on a thumbnail registers. So does the gap in seconds before you hit skip, the clip you exit at once, the remark you leave behind, the phrase you type into search. One of those on its own is noise. Stack them across hundreds of hours spent on the platform, though, and they resolve into an unexpectedly faithful portrait of what you like.

The trick is not unique to YouTube; plenty of AI systems lean on the same one. Anyone who has read our explainer on how AI picks out faces in photographs will recognise the shape of it: an unruly, very human thing — a face, in that instance — gets turned into a tidy set of numbers suitable for comparison and measurement. Viewing behaviour receives identical treatment here. Viewer and video alike are squeezed into a long sequence of numbers known as an "embedding," effectively a mathematical fingerprint of content and of taste.

Similar items are deliberately placed near one another inside a vast mathematical space built for the purpose. Videos about espresso machines gather in one region, and anyone who has been through a dozen of them acquires an embedding nearby. Recommendation then becomes a matter of hunting for the embeddings closest to your own. Should the idea of "training" a system to behave this way be unfamiliar, the fundamentals are set out plainly in our guide to machine learning and how training works.

02From App Launch to Feed: YouTube's Decision Path

Here is the precise route taken between opening the app and being handed a feed:

03Try It: Score a Video the Way the Algorithm Would

The relative weight of each signal is worth poking at. Push the buttons underneath and watch a video's score swing as different kinds of viewing behaviour get applied, all through a deliberately reduced stand-in for how YouTube orders videos.

04What Sits Underneath: Neural Networks and Transformers

Beneath the ranking stage lies a deep learning architecture commonly called a "two-tower" model. The first tower derives a numerical portrait of the viewer from that person's history. The second derives a numerical portrait of the video from its title, description, thumbnail, audio, and the reactions of other viewers. Training pushes these two towers nearer to each other in mathematical space whenever a genuine viewer genuinely enjoyed a genuine video.

Newer iterations of YouTube's recommendation stack have moved toward transformer-based architectures as well. That is the same family of model sitting behind modern chatbots and translation tools, and it lets the system read your history as an ordered sequence rather than an unordered heap. Order and context are what a transformer pays attention to — much as it tracks word order within a sentence — instead of lumping your last fifty videos into a bag. Our guide to what a transformer model actually is digs into the mechanics of that attention mechanism.

A tidy parallel exists with large language models. As our explainer on how AI settles on its next word describes, such a model forecasts the word that follows, given everything preceding it. YouTube's sequence models do the structural equivalent: forecasting the next video, given the run of videos already behind you — trained on viewing patterns rather than on language. None of it functions without colossal quantities of training data, a point our article on why AI training demands so much data unpacks, since YouTube's models learn from signals produced by billions of watch sessions every day.

05Every Corner of YouTube Run by Recommendations

Claiming this is merely one feature undersells it badly. Beneath almost every surface the platform presents, the same engine is at work.

06Just How Accurate Is It — and Where Does It Slip?

Predicting your clicks and your staying power is something this system does unusually well — that being the metric it was tuned against — and the results are visible. Being skilled at predicting clicks, though, is not the same as being reliably good for you, and there are unmistakable gaps in what the model can see.

Where the Algorithm Comes Unstuck:

1. ✗

   The Cold-Start Problem

   ✗ A newcomer and a just-uploaded video bring no history for the model to work with, leaving early suggestions as crude approximations drawn from general popularity instead of your particular taste.

2. ✗

   Living in a Filter Bubble

   ✗ Since every suggestion echoes something you have already watched, the feed can narrow into an echo chamber without you noticing — unless you deliberately go looking for unfamiliar subjects.

3. ✗

   The Pull of Clickbait

   ✗ Overblown thumbnails and titles can lift click-through rate for a while even when the video underneath fails to deliver, tilting the short-term signals before watch-time figures pull things back.

4. ✗

   Interests That Have Moved On

   ✗ When your interests turn over faster than your history does, the model carries on pushing an abandoned hobby long after you have drifted to something else.

5. ✗

   One Account, Several People

   ✗ A single account shared across a household generates tangled and frequently contradictory signals, because the model is simultaneously attempting to please several different people.

07Privacy and Ethics: The Unsettled Questions

When prediction of human behaviour reaches this standard, the debate widens considerably — accuracy of any individual feed stops being the whole question:

Control has not been taken away entirely. Pausing or clearing your watch history, marking something as "Not interested," or choosing "Don't recommend this channel" — each of these pushes fresh, intentional signals into the system, steering the algorithm somewhere you would rather it went.

08Questions Readers Ask Most