Every second you watch a TikTok is a vote, and the app counts the votes continuously. Watch time, rewatches, and how fast you skip all feed straight back into the model that picks your next video. You are not just using the algorithm; you are training it, with almost every moment of attention you give it. And it measures attention, not approval, which is where things get interesting.

This is the single most useful thing to understand about how the For You page learns you so fast.

Attention is the signal

When the system tries to predict whether you will engage with a video, the best evidence it has is what you did with similar videos before. And the richest, most reliable piece of that evidence is time. How long did you watch? Did you make it to the end? Did you let it loop and watch again?

Time is a great signal because you produce it constantly and you cannot easily fake it. A like is one bit of information and you might give it for politeness. A thirty-second watch is thirty seconds of demonstrated holding power. Behavior beats opinion, and watch time is behavior in its purest form.

So the model leans on it heavily. Lingering says "more of this." Quick skips say "less of this." Rewatching says "much more of this." You are issuing instructions with your eyes.

It measures what holds you, not what you like

Here is the part worth slowing down for. The system measures attention, not approval. Those feel like the same thing. They are not.

A video can hold your attention because it delights you. It can also hold your attention because it annoys you, worries you, confuses you, or makes you want to argue. From the outside, all of these look identical: you stayed.

This is why hate-watching backfires. You watch the thing that makes you roll your eyes, maybe you watch it twice to be sure it is as bad as you think, and the model dutifully records a long, repeated watch. It has no way to know you disliked it. It only knows it worked. So it shows you more of the same, and a whole genre of content you do not even enjoy can quietly colonize your feed.

The feed is not a mirror of your taste. It is a mirror of your attention, and the two are not the same.

You are training it whether you mean to or not

You might think the algorithm learns mainly from your taps, your likes, your follows. Those help, but they are the small part. The large part is passive. Just by watching, you emit a steady stream of signal the whole time the app is open.

That means there is no neutral, observer mode. There is no "just looking." Every session is also a training session. The longer you are in the feed, the more you teach it, and the better it gets at keeping you there. This is part of why the loop tightens over time and why the app can feel like it is closing in on exactly what keeps you watching.

It also explains why the feed feels relentless once you have fed it heavily. You built that. Not out of weakness, but simply by being present and attentive, which is the one thing the design is engineered to harvest.

Turning the signal around

The good news is that the same mechanism runs in both directions. If watching trains it, then watching on purpose is a way to retrain it.

A few concrete moves:

  • Linger on what you actually want more of. Give your genuine interests real watch time so they outweigh the accidental hooks.
  • Swipe away fast from the stuff you don't want, before a long watch registers as a vote for it. Speed is the signal here.
  • Use "not interested" when it's offered. An explicit thumbs-down is a cleaner instruction than a fast skip and the system weighs it directly.
  • Stop hate-watching. It is the most common way people accidentally poison their own feed. If something makes you angry, the most effective response is to leave it unwatched.

None of this works instantly, because the model is balancing many guesses about you at once. But over days, consistent signals do steer it.

The deeper lesson is the calm one. The algorithm is not reading your mind or judging your character. It is counting seconds. Once you know that watching is voting, you can vote more carefully, and you can simply give it fewer seconds to count. The settings that slow the feed are the practical extension of exactly this idea.