Any software engineer worth their salt could tell you the immense difficulty of capturing analogue data, modelling it, and translating that model via a set of rules into an action. And repeating that continuously in real time. The more variables and ambiguity that are present in the input, the more likely it is to screw up in the output. In a 2 ton vehicle going at 70mph along a road with other traffic, that could be positively fatal.
I wouldn't trust any autonomous mode unless it requires an alert and attentive human being at the wheel and forces their attention. At least that way the human can veto or correct the car's actions.
This isn't exclusive to Tesla. Any autonomous vehicle that claims it can drive itself in limited, or unlimited circumstances still requires oversight. Otherwise it will do something dumb and/or dangerous and there will be no human paying attention to stop it.