Clearly, you want your list of "frequently searched terms" stored locally on your device in a very small and efficient history file.
However, if you store this file on the server, you can hide from the user what actually gets stored in it, it takes longer so it seems like it's doing harder work, and for some reason gets it wrong occasionally which means -- AI.
Google Maps is like this. It completely refuses to remember your recently searched addresses if you disable Location History (which includes remembering and storing, let's call it a little bit more info than just my recent search terms).
This would be such a prime candidate for storing securely, privately on your device, for any type of map service, that I can only conclude this is deliberate hostile anti-user programming.
Also I bet there's code out there that just returns most-recently-searched with a few deliberate mistakes to seem more opaque and thus more AIey.
I'm sure there are companies which play the various games you're suggesting, but I think positing that it's the rule verges into the conspiracy theoretical.
For sure, with all the type of Neural Structured Learning, seems like we are just trying things out by training models. Would be good to have a way to actually explain to us developers how decisions are actually made. I know it is based on some kind of statistics.
If anyone can point in the direction that would be greatly appreciated.
The sorts of things phones used to remember I'd never refer to as AI. A list of contacts sorted by the frequency with which I use them isn't AI. A list of contacts sorted in an order I don't understand, with a slight preference for frequent contacts, is.
The latter have proliferated recently, hence the shift.