Kuaishou's New AI Knows Why You Like a Video

Today's recommendation systems — the software that decides what videos, songs, and posts you see — mostly study your behavior. They track what you click, how long you watch, and what you skip. The problem, as the researchers put it, is that these signals reveal what you do, not why you do it. You might watch a cooking video because you love cooking, or because you're hate-watching it. The algorithm can't tell the difference.

A team at Kuaishou, a short-video and live-streaming app with hundreds of millions of users in China, decided to collect the "why." They built a system called SARA that asks people to explain their preferences in plain language — "great host," "too much shouting," "funny and useful." They collected these explanations from 240 million users. But most people never write anything, so they trained an AI to write plausible reasons on their behalf, expanding coverage from about 86,000 people to 10 million creators. That AI then feeds those reasons into the ranking system that orders your feed.

In live tests run for over a month, the upgraded feed kept people watching longer and — importantly — reduced negative feedback, meaning fewer people slammed the "don't show me this" button. That's the rare double win: more of what you enjoy, less of what annoys you.

The catch is privacy and sparsity. The system works best when people volunteer honest opinions, which few do, so much of it is AI-inferred guesswork that can be wrong. It also means platforms are storing and modeling your stated tastes, not just your habits. And so far it's only proven inside one Chinese app.