New AI Trick Makes Short-Video Feeds Show More Fresh Creators

Here's a problem you've probably felt: your short-video feed keeps serving up the same viral clips and the same big creators. That's not bad luck — it's a feedback loop. When an app's AI recommends popular videos, it collects a mountain of data about popular videos. It then learns to spot 'popular' patterns extremely well and everything else poorly. The huge catalog of smaller, newer, niche creators — researchers call it the 'tail' — gets quietly ignored.

The fix, from a team publishing at the ACM RecSys '26 conference, is a gentler training method they call 'soft curriculum learning.' Think of it like a tutor who starts you on easy problems, then gradually adds harder ones instead of throwing away the tough questions entirely. Older approaches did throw data away, but that wasted expensive AI chips and slowed everything down. This method instead turns the difficulty up and down with simple math adjustments, so the hardware keeps humming along at full speed.

In live A/B tests — real users, real app, split into two groups — the platform saw meaningful gains in both overall viewer satisfaction and how much fresh content people watched. The researchers tested the approach across three kinds of recommendation engines, including ones that predict what you'll want next based on your viewing history, and ones that match you to videos in two steps. All improved. Crucially, none of it slowed the system down, which is usually the dealbreaker for this kind of change.

So what does this mean for you? A feed that surfaces new creators is a feed that feels less stale, and it gives smaller voices a genuine shot at being discovered instead of being buried under the algorithm's favorites. It's a research paper built for giant platforms, not a switch you can flip today, so expect gradual changes rather than an overnight overhaul. But the direction is clear: your 'for you' page may soon actually be for you, not just for what everyone else already watched.