When cheap AI videos outnumber human ones that viewers like more, does the feed widen that gap or close it?

This explores whether platform recommendation loops widen or close the gap between what creators supply and what audiences actually want. The clearest case in the collection is AI-generated video. On a Chinese short-video platform, viewers watched AI videos less often and less completely than human-made ones. AI creators still earned about the same total engagement, because making each video costs them so little that they simply post more of them Can AI creators match human creators through posting volume alone?. That's the imbalance in concrete form: volume can make up for weaker preference. The real question is whether the feed rewards volume or corrects for it.

The default behavior of ranking systems points toward amplification. YouTube's ranking work shows that a model trained on its own past recommendations learns from clicks that its earlier choices caused. Unless it explicitly corrects for that, for example by accounting for where an item sat on the page, it drifts toward a degenerate equilibrium that reinforces whatever it already favored Why do ranking systems need to model selection bias explicitly?. Creators respond to that too. Feed weights change what producers choose to make, so the loop runs through creator behavior as well as viewer behavior How do recommendation feeds shape what people see and believe?. A high-volume creator who gets early exposure therefore produces data that justifies more exposure, and other creators adjust to match.

There is also evidence of dampening. On the same platform, the algorithm gave AI-generated content less exposure than comparable human content across nearly 180,000 matched pairs. That could offset the advantage from sheer volume Can algorithmic distribution prevent AI content from overwhelming creator diversity?. One caveat: the excerpt in the collection reports the direction of this effect but not the full exposure results or robustness checks, so treat it as suggestive. The broader point is that a feed isn't neutral plumbing. Its designers can tune it to push against imbalances, but that takes a deliberate choice; it doesn't happen on its own.

The less obvious finding is that the type of loop matters as much as whether one exists. Product recommendations built on "frequently bought together" and on "also viewed" produce opposite effects on ratings. One pulls opinions together and the other pushes them apart, because each brings a different audience with different expectations to the same products Do different recommender types shape opinion convergence differently?. A related warning comes from AI training. Giving each user their own reward model removes the averaging that held extremes in check, and the system drifts toward flattery and echo chambers Does personalizing reward models amplify user echo chambers?. The more finely a loop is tuned to individual signals, the less it corrects anything.

There's also a cost that engagement numbers miss. Even when AI posts win on engagement, the credibility and attention they collect don't build up any lasting reputation for a real person. Over time this erodes social media's role in surfacing trustworthy human voices Does AI content displace human influencers on social media?. So an imbalance can look balanced on engagement metrics while it quietly damages something the metrics don't measure. The collection doesn't yet contain long-run studies that track whether this kind of exposure adjustment holds up as AI's share of content grows.