This is a pretty hand-wavey analysis. I believe there is rich potential for a rigorous, formal, statistical analysis of "good taste", that no one in the literature has done yet.
I will outline a few ways that one can formalize "good taste". I understand that there are some weaknesses and gaps in the concepts that I will delineate. This isn't because I think my analysis is final. Rather, I am suggesting the starting point for this sort of inquiry, which through refinement by other researchers could actually become stronger as a field of inquiry.
There is an economy of attention, and any analysis of taste or preference should be based upon how one spends their attention.
One argument might be that a particular kind of good taste is being able to anticipate what someone else will like. This is a demonstrable skill that some people (and recommender systems) possess, and others don't. There might be other kinds of good taste besides prediction, but this is one important component that can be measured.
Important confounding variables is bias caused by other people. For example, no one likes a particular artist until a famous critic pronounces them as good. This is a widespread confounding variable, but nonetheless could be avoided in certain controlled experimental setups. Again, this isn't helpful when we are talking about quantifying taste in the real world, which bias is unrestricted.
Another form of statistical analysis would be to say that people with broad undifferentiated preferences ("pop") have less refined taste than subgroups with niche specialized taste. Possible analyses here include: Are there subgroups with refined taste that is not just associated with a specific subgenre, but extends across many genres of this particular medium? That suggests a broader sort of refined art taste than generalizes and isn't based just upon some expertise. Additionally, detecting people with "random" taste that isn't correlated with the taste of other people suggests the person is just throwing darts and being contrarian, not that they have some taste that suggests a deeper shared human understanding.
One weakness of existing recommender systems is that like/dislike and five star rating systems rarely quantify: "Wow this is so amazing I would sacrifice my right arm for this." This is because there is no economy of "five stars" ratings in most systems, and the number of five star items is potentially a large percentage of the whole corpus of art. Instead, a Michelin-star like system could zoom in on the 1% of art that has a really transformative impact on the listener.
About the objection pg says that: "Well, we might think some artist sucks now but in one hundred years they are revered." I think this argument can also be refuted. Within the context of art analysis in the 1800s, a particular artist might make no sense, because their work is too prescient. Whereas within the context of later artists who allow the public to appreciate the work of the dead artist, liking the dead artist now contextually becomes good taste.
Again, I don't think I've presented a conclusive or bulletproof analysis here. I've just tried to outline how a formal and rigorous approach to quantifying "good taste" is an endeavor we could actually perform and engage in, but I haven't really seen in my review of the literate yet. There might be some important works that I've missed, perhaps in machine learning philosophy. It's easy to hand wave through saying "good taste" doesn't exist, but I think there's value in challenging that assumption and seeing how far we can get at formalizing it, and what potentially illusively remains nebulous and is actually bullshit.