One feature a few of us toyed with for years that I'd love to see tried seriously someday (or does it already exist?) would be intelligent tag trees or maps that stuff could be plugged into, ideally that could also be shared and mixed too. More specific tags would have parent/child/sibling relationships with more generic tags, and a single tag would then pull in all the others as far as searching. So for example if I have a photo of one of my cats, and I tag it Cat Name having set that up, "calico" or "tabby", breed, plain old "cat" and anything else associated with the specific tag would all be hits. And up the stack, if I tag a picture "calico" it should show up under the "cat" tag as well automatically with no further effort. Conditionals could be incorporated too, like if there is a character who in a series has long hair, tagging "short hair" could automatically pull in "alternative hairstyle".
My big problem with typical tagging is that it's a lot of grunt work to be comprehensive. AI can help but has its own issues. While getting the UI/UX right would take some work, a universal system-wide multi-dimensional tag map could allow a lot more metadata a lot more easily. Even more so with sharing, communities could curate tag maps around areas of interest and one could stay synced with tag maps of "aircraft" or "trees" or the like which would bring in lots of technical depth to explore in a distributed fashion. Eventually maybe community AI recognition models could be an optional part of that too to help identify specific given subjects (and one would be free to then not have other kinds of things identified or shared).