Good and accurate points. Human drivers often use the follow-the-leader strategy too, especially on unfamiliar roads. What cues would a typical human driver have picked up on to cue them to danger in this situation? The assumption is that the driver would have visually seen the crash barrier ahead, or if not, would have at least seen the pattern of lanes splitting off and deduced that the present "lane" was not really a lane. Tesla's current system seems to have far too much tunnel vision when it comes to lane identification.
There are all sorts of cues, likely not within AP's information repertoire, that would have signaled to a human driver that there was likely to be a gore to the left of the fast lane and would have indicated to a human driver that keeping relatively to the right was a good driving strategy for staying in lane:
1) A human driver would see the overhead signs announcing that there was a left side exit, and exits mean gores.
2) A human driver would see that the cars in the exit (HOV) lane were separating left relative to his lane, rather than maintaining a parallel path.
3) A human driver would see that the exit (HOV) lane became a fairly steep and visible ramp ahead.
4) A human driver would get a gist of the general direction of traffic in the fast lane, and in other lanes of the mainline of the highway.
All of these could cue a human driver that the fast lane was keeping left-- regardless of the behavior of the specific car the driver was following.
These cues would have been much stronger if the human driver mistakenly did enter the gore, and would have been strongly supplemented by (i) seeing the concrete barrier at the end of the gore and (ii) seeing that the distance between the "lane lines" (gore edges) was widening, like a cone, rather than staying constant, as would be normal in a fast lane. I doubt it is common for a human driver, once in a gore, to travel all the way to the concrete structure at the end of the gore without ever realizing he or she is in a gore. yet it seems like AP, once it found itself in the gore, was incapable of realizing that something was fishy and it was in a gore rather than in an oddly shaped lane.
More importantly, most drivers drive the vast majority of their driving miles on a relatively small part of the road system. So most of the time a human driver isn't driving a road that is "new" to them. Most of the time in fact humans are driving in lanes they have driven many many times before. A human driver is very unlikely to accidentally move into a gore on a roadway they are familiar with. They might enter a gore to make a last minute (and dangerous) lane change; but they are unlikely to enter a gore on a familiar route because they confuse the gore with the traffic lane.
AP seems to lack this ability to learn the "lay of the road" from repeated trips over a road. And it doesn't seem to really use maps much (if at all) to simulate that knowledge. Therefore, every road is a "new" road to AP on every trip. This difference between humans and AP is especially stark in this instance, where this road was part of the driver's commute route and the human driver (and the specific car) had probably been over this exact spot countless times.