TLDR if you don't wanna watch the whole thing: Benaminute (the Youtuber here) creates a fresh YouTube account and watches all recommended shorts without skipping. They repeat this 5 times, where they change their location to a random city in the US.
Below is the number of shorts after which alt-right content was recommended. Left wing/liberal content was never recommended first.
- Houston: 88 shorts
- Chicago: 98 shorts
- Atlanta: 109 shorts
- NYC: 247 shorts
- San Fransisco: never (Benaminute stopped after 250 shorts)
There however, was a certain pattern to this. First, non-political shorts were recommended. After that, AI Jesus shorts started to be recommended (with either AI Jesus talking to you, or an AI narrator narrating verses from the Bible). After this, non-political shorts by alt-right personalities (Jordan Peterson, Joe Rogan, Ben Shapiro, etc.) started to be recommended. Finally, explicitly alt-right shorts started to be recommended.
What I personally found both disturbing and kinda hilarious was in the case of Chicago. The non-political content in the beginning was a lot of Gen Alpha brainrot. Benaminute said that this seemed to be the norm for Chicago, as they had observed this in another similar experiment (which dealt with long-form content instead of shorts). After some shorts, there came a short where AI Gru (the main character from Despicable Me) was telling you to vote for Trump. He was going on about how voting for "Kamilia" would lose you "10000 rizz", and how voting for Trump would get you "1 million rizz".
In the end, Benaminute along with Miniminuteman propose a hypothesis trying to explain this phenomenon. They propose that alt-right content might be inciting more emotion, thus ranking high up in the algorithm. They say the algorithm isn't necessarily left wing or right wing, but that alt-right wingers have understood the methodology of how to capture and grow their audience better.
Even in the best-intentioned recommender system, trained on the content you watch to estimate what you're interested in and recommend similar things, that would be the drift of things. You can't really mathematically judge the emotions the viewers might feel unless they express them in a measurable way, so observing their behaviour and recommending similar by whatever heuristic. And if they keep clicking on rageposts, that's what the system has to go on.
But at least giving the explicit indication "I don't want to see this" should be heavily weighted in that calculation. Just straight up ignoring that is an extra layer of awful.