This report has a really, really critical flaw:
The latter I refer to as “shadow mode” because the data streams in these vehicles (whether under manual or Autopilot control) is available to be used for training the neural networks that perform the various components of the perception-control task.
The problem is that
the cars aren't transmitting data.
You can hook up any Model S/X/3 to a WiFi network and monitor the router usage. The payloads it's uploading are quite small (largest I saw was 20 MB), certainly not large enough to be either compressed or uncompressed video data, and especially not for 8 separate cameras. I can't test for 4G connections, but I doubt that Tesla would be using this to stream video, as it would get quite expensive.
Since the most challenging part of autonomous driving is rapid object identification from video feeds, it doesn't seem that Tesla cars are currently "uploading data" in any meaningful way.