If something like that were to work, a lot of effort would need to be put into minimizing the UI friction. I could see something like: uploaders add topic tags to their videos, and an AI runs in the background to generate and apply new tags based on the content (most people would not understand how to properly tag content). An AI would also be used to create a graph of related tags, where similar or closely related tags are nodes joined by an edge. Then, on first login the user is prompted to pick some tags to start with. Over time, the client uses the adjacent tag graph to fine-tune users’ tags, on device. The idea here is that we could get a decent algorithm that can recommend new stuff based on what the user watches, but keep that data processing of user-specific content local. Then, the client would also have an option the user could enable that would contribute their client’s tag information back to the global tag graph, improving the global tag graph for everybody. This data could also be combined with other users data at the instance level to somewhat anonymize the data, assuming it is a large multi-user instance. If you were to host a single user instance, you’d probably not want to contribute to the global tag graph unless you’re ok with your tag preferences being public.

It’s a bit tricky but I think a privacy preserving algorithm is possible. Simply put, the more data available, the better an algorithm can be.