I think it will be like TACC, Initially you just test it, wait for the edge cases where it fails. This is how you will determine what the car can or cannot manage. How would you describe the limitations of TACC into words everyone will understand without trying it? IMO you couldnt and even if you did it would take a rather foolish person to put blind trust in their understanding of said text, without testing it personally.
The difference here is that, unless there's a radar or camera malfunction/blockage, TACC can ALWAYS be engaged at speeds 18mph or above, regardless of what the road conditions are.
With Autopilot (TACC + Autosteer), the road conditions and location DO matter, which means sometimes I may get the full Autopilot (including Autosteer) and sometimes I may just get TACC. This will depend on whether the road is "relatively simple" or not.
So, to even get Autosteer to engage, it sounds from Elon's comment that the road must be "relatively simple". If it's not, Autosteer won't engage, but TACC will. It would be nice if, as a driver, I know ahead of time what "relatively simple" means to the car so I know what to expect.
Of course, whether or not Autosteer is engaged could be clearly communicated to the driver at the time Autopilot is engaged, but then there must also be a very obvious warning when Autopilot is engaged but the car determines the road is no longer "relatively simple" so the driver can take over the steering function.
My point is, if I know as the driver what those conditions are, I would know (even potentially ahead of the car) when to take steering back over and not have to wait for the car to tell me to.