It seems fairly obvious me that an imperfect autopilot is worse than no autopilot at all, since it trains drivers to “relax” and let the computer drive when some non-trivial percentage of the time the autopilot will get things wrong.

It’s unfortunate that there’s no real way for there to be a perfect computer driver, and all the real world training data these things are collecting is needed to ever get close.

But if these systems keep making mistakes like this that seem “stupid” to humans, I suspect these things will be banned before long.