Of those 3 ways, ML would likely be the most consistent, though its learning.
Community consists of people viewing criteria differently, meaning that you can literally do the exact same submission twice. Get a random sample of the community or thinks against the submission and fail, and get a random sample of the community that thinks for the submission and pass. Heck, even a submission that overall majority community thinks should pass or fail, can get the opposite result by getting a lucky/unlucky oversampling of the minority view.
Even Niantic’s team is human. You will find many results posted on here if things accepted/rejected by Nianatic Reviewers that the vast majority of the community disagree with.
One last thing worth repointing out, is that in game POIs aren’t always the best example. Some of them are older than the existing standards and review system. Some of them got a lucky subset of reviewers when the majority of the community would have rejected. That tends to frustrate submitters the most as they often say “Well I’ve seen X”
See to me memorial benches rarely are sensitive. That being said the hospital campus I work at has several memorial benches that are along the outdoor walking path, so often those are what I think of when talking about memorial benches. I guess it’s all about context. For example the chapel at work has a lone dinner table all set up inside the chapel amongst the pews. Is it a strong memorial and possibly in some ways a great wayspot? Probably, but it’s for certain in what I would call a sensitive area I would not want people gathering with their nose in their phone, so I personally have never submitted it ( and I rejected it the one time I saw it ). The outdoor memorial benches were wayspots before I started working here. Honestly I will say the wayspots around campus actually promote exercise because each one isn’t that far away, but if I just go to the next one I end up walking about a mile around campus.