So, am I correct to understand that the recommender AI is just learning from youtube users' actual behavior and shows what content people who have watched a given set of videos tend to find "engaging" (watch for a significant amount of time, leave comments, like, etc)? That, in effect, the AI is just holding a mirror to youtube users, and some journalists don't like what they see in it?
There is a known problem with Youtube both excessive recommending in the same category and driving people to more extreme videos.
If you watch a surfing video, even for a few seconds, expect a lot of surfing videos to pop up in your recommendations for a long time. This has gotten slightly better over time in my experience.
The larger problem is extremism. Watch a video about a vegetarian dish, get recommended vegetarian dish and lifestyle videos, watch those, get recommended veganism videos, watch those, get recommended PETA and animal activism. It kinda works like a funnel.
Might not seems that dangerous for your favourite topic, but now replace the topic with conspiracy theories, religious extermism, radical feminism, anti-government armed resistance, self-harm and suicide, anorexia etc. etc.
All of this reminds me of the early days of the TiVo. It had a recommendation system that it would use to fill your DVR with shows you would like. If you recorded a show in a new category (which gave the show an implicit positive rating), it would often go off the rails and find everything in that category. The first time I recorded "The Simpsons", it filled up my DVR with cartoons over the course of the next week before I noticed what it was doing.
We seem to be basically no better than that sort of logic.
> There is a known problem with Youtube both excessive recommending in the same category and driving people to more extreme videos.
There is a belief that that problem exists with YouTube but whether that actually exists requires further investigation. For example [1] suggests the opposite.
a) we don't know how the recommender works, and it changes over time, and some problematic recommendations have been found even in absence of actively seeking out problematic content, and
b) it's not journalists but researchers from mozilla and scientists from the University of Exeter.
> some problematic recommendations have been found even in absence of actively seeking out problematic content
I don't know how to put it in proper terms, but isn't it possible that the user who has watched videos A and B ("non-problematic") is classified the same as a subset of users who have watched A, B and C (C being "problematic"), and thus is shown C?
Asking because personally, I've never seen my "safe for work" youtube account ever recommend me anything "problematic". At the moment, it's just feeding me music videos of the same genre, which I am already bored by :-)
Some studies such as [1] measure recommendations without past browsing history, and they still find recommendations for harmful content.
Other studies (such as this from Berkman Klein Center [2]) simulate user behavior and subsequent recommendations and find additional harmful effects of personalized recommendations.
[1] Faddoul, M., Chaslot, G., & Farid, H. (2020). A Longitudinal Analysis of YouTube’s Promotion of Conspiracy Videos. ArXiv:2003.03318 [Cs]. http://arxiv.org/abs/2003.03318
So, am I correct to understand that the recommender AI is just learning from youtube users' actual behavior and shows what content people who have watched a given set of videos tend to find "engaging" (watch for a significant amount of time, leave comments, like, etc)? That, in effect, the AI is just holding a mirror to youtube users, and some journalists don't like what they see in it?