The problem is that:
1) It's very difficult to manually or automatically scrub identifying information from video, and doing so probably greatly reduces the value of the data for AI training purposes
2) If you allow users to delete videos that have been sent, then you no longer have a reproducible corpus of data
3) If you keep the recordings for say, a week, and allow the user to delete them before they're sent, either users don't have the time to review it (reviewing hundreds of Sentry clips comes to mind), or are going to just forget about it and we'll have the same problem.
I suspect they do not want to talk about the data collection too much in general, because just like how Spotify built a major value add (we have lots of playlists) off of tricking users into making user playlists public by default (it's actually very non obvious when playlists are created), having a huge fleet of cars recording video is hugely valuable for development of self driving.
I think where I'm conflicted is that at the same time as I personally consider this kind of thing an invasion of privacy and overall a nuisance, I recognize that there is societal good/value to be gleaned from mass collection of datasets. If we asked users whether they wanted to share their playlists or car video recordings, you would end up with almost no data, and the data you got would invariably be biased (see Apple Music and the paucity of good user playlists).
Clearly there's a missing middle solution, but I'm not sure there is a good one.