Feeds converge quickly on a narrow band of material, and the narrowing is not a fault in the system. It is the system working as designed.
The system is predicting one thing
A recommender estimates the probability that a given person will engage with a given item, then ranks by that estimate.
It has no concept of variety, quality or whether the viewer would prefer something new, because none of those are the quantity being estimated.
Anything that improves prediction accuracy is therefore adopted, and similarity to past behaviour is the strongest predictor available.
The signals used are also cruder than viewers assume. Watch time and whether a video was replayed carry more weight than an explicit rating, because far more people supply them.
Certainty beats potential
Showing something similar to what a viewer already watched carries a high expected value with low variance.
Showing something unfamiliar might produce a much better outcome or a much worse one, and the ranking has no reason to accept that spread.
Repeated across a session, the safe choice wins every time, and the feed contracts without any decision to contract it.
The feedback loop closes on itself
The system learns from what a viewer engages with, but a viewer can only engage with what they were shown.
Material never surfaced generates no signal, so the absence of interest is indistinguishable from the absence of opportunity.
Each round reinforces the previous one, and after enough rounds the model has a confident picture assembled from its own recommendations.
A single unusual session can therefore reshape a feed for weeks, because the system cannot distinguish a passing curiosity from a new interest until it has tested the idea repeatedly.
Exploration exists but is rationed
Platforms know this and deliberately inject unfamiliar material to keep learning about a viewer.
That exploration costs engagement in the short term, so the proportion is kept small and is often reduced for viewers who are already highly active.
The people with the most established habits therefore receive the least variation, which is the opposite of what they usually report wanting.
Why creators converge as well
Creators observe which of their own posts were distributed widely and produce more of that, because distribution is the entire economy they operate in.
Formats that the system rewards spread across everyone making material for it, and the supply narrows alongside the demand.
The sameness a viewer notices is therefore produced twice over, once in what is recommended and once in what was made to be recommended.