This is what I was thinking. If you stop a second to think about it from their perspective, it can make more sense. (I know in 2020 it is unheard of to think about something from a different point of view than one's own.) Even real-life human beings in the assistant role get things wrong as they interpret the original request not as intended.

In that case maybe the assistants should be more like children and like ask questions?

Relatedly, I wonder if these assistants “filter bubble” you like search does. Like, learning what types of things you are looking for and grouping you with other similar people.

I believe current evidence is that children grow both intelligence and knowledge into adolescence. After that, it's mostly knowledge. i.e. both compute and data grows for quite some duration, then it's mostly data.