Former USPTO patent examiner here. Patent similarity search doesn't work that well at present. Even when there is a perfect 102 reference (the best sort of prior art reference; the reference shows exactly what is claimed) in the patent literature [0], frequently none of the 3 similarity-based search engines I'd try [1] would return that reference. There are many reasons for this. In my previous area (water heaters and ventilation, mostly), a lot of the issue was because the software would only look at the text (and citations) and the drawings really did provide a lot of critical details that aren't easy to express in text. (I already would check the provided citations so the AI checking the citations is not a value-add in my opinion.)
Patent searching is not anywhere near as easy as many people here seem to think! That's particularly true when the examiner has a very limited amount of time to search, which is every time.
[0] Yes, searching the entire web in addition to patents would be better. But it's already clear that the software often doesn't return the best references from the patent databases it does search, so I'm skeptical that moving to the entire web would make that much of a difference.
[1] IP.com (can work on unpublished applications), Dialog Foreign Patent Finder (can work on unpublished applications), Google Patents (if the application was already published). Each of these seem to be roughly as good as each other. For a while I'd also try the USPTO's home-grown PLUS system from the early 00s, which returned mostly irrelevant documents.