YouTube's New AI Can Tell If You Actually Liked That Short
Your watch time was lying to YouTube — now it can sense real enjoyment.
Every recommendation feed — YouTube Shorts, TikTok, Instagram Reels — tries to guess what you like by watching what you do. The main clue is watch time. But researchers point out that watch time is a bad lie detector. Watching a video all the way through might mean you loved it, or that it was only eight seconds long, or that you zoned out with your thumb hovering. Two very different feelings look identical in the data. Worse, when companies optimize for raw watch time, they accidentally reward short, clickbaity clips over ones you'd actually be glad you saw.
The fix, described in a paper presented at the RecSys 2026 conference, is a model the authors call FLVM. Instead of treating your behavior as a direct answer, it treats it as a fuzzy clue pointing to a hidden quality — how much you really valued that video. The system splits the work in two. One part accounts for boring, predictable factors like how long the video is, how prone you are to binge-watching, and what time of day it is. The other part hunts for the genuine spark of interest. That score then slots into the existing recommendation system as one more signal for ranking what to show you next.
The team tested it on YouTube Shorts, one of the biggest short-video platforms in the world. In live A/B tests — where some viewers got the new system and others didn't — a primary measure of viewer enjoyment rose 2.67%. That sounds small, but at YouTube's scale a couple of percentage points touches hundreds of millions of people every day. The researchers also say it improved their offline quality metrics, meaning the model looked better on historical data too, not just in the live trial.
The honest caveats: this is a technical paper, not a product launch, and 'viewer enjoyment' is the company's own internal metric, which outsiders can't inspect. It also doesn't promise less time in the app — better recommendations could just as easily mean you stay longer. And the researchers don't claim to solve the deeper question of whether endless short videos are good for us at all.
- Recommendation algorithms mistake watch time for enjoyment — a short clip you barely noticed can score as high as a video you loved.
- The new model, tested on YouTube Shorts, uncovers the hidden 'did you really like this?' signal and raised a viewer-enjoyment measure by 2.67%.
- Expect your feed to feel more tuned to your actual taste, though it doesn't guarantee you'll spend less time scrolling.
Why It Matters
Better recommendations could mean less doomscrolling and more videos you're genuinely glad you watched.