My concern is not that DE's do ETL, just that a data scientist could actually avoid learning how to do this.

Certainly in all the successful orgs I've worked in, this would rarely be the case. Maybe it works if you have really well-standardized and static data, but I rarely work in this kind of environment.