Related to what
@BerTX just posted, I'd say it's part myth and part fact. There's a perceptual illusion where we tend to think another vehicle is getting closer as we pass by. Without AP we may subconsciously give it more room in the lane, and anyway we're probably keeping most of our attention on the road ahead — I hope. But AP tends to stay centered in the lane, and with AP we have more leisure and more cognitive bandwidth to notice these details. I think that accounts for
some of the anecdotal reports of "truck lust". We could call these "category 1". But on the face of it we
can't push all the reports into that first category.
I've seen AP HW1 apparently mistake a shadow or a guard rail for a lane line. If that can happen, then it's easy to see how it might mistake part of a long trailer for a lane line. The car may swerve suddenly to try to center itself in what it thinks are the new lane lines. Call these reports "category 2". An abrupt swerve like that wouldn't be unreasonable if the lane lines really did change. Construction zones often have abrupt changes like that, and I drive past a couple of more permanent examples on a regular basis. In those locations drivers should swerve in a fairly abrupt way — if they're paying attention and following the lines.
If all this is correct, category #1 may be worrying but should be safe to ignore. Category #2 is an error and potentially dangerous. Can it be fixed? Maybe, but not necessarily. Keep in mind that the AP HW1 camera is very limited. It's like a one-eyed person, wearing blinders and staring at a fixed point not too far in front of the hood. Tesla (and Mobileye) did an amazing job with that limited vision, but software can only do so much when the input data is poor. Say the camera and its machine-learning backend can't be taught to distinguish between a real, abrupt lane change and a truck-induced artifact. Then the programmers are left with a choice between failing on the real, abrupt lane change and failing on the truck-induced artifact. Neither failure would be safe, but they may have enough data to decide which one is likely to be safer.
Limitations like that throw more of the burden back onto us, the drivers. We have to understand the system's weaknesses and supervise it closely.
HW2 has potential to be much better: three front cameras, each with a different focal length and angle of view. That should give the software enough information to avoid some, perhaps all, of these errors.