The problem is that the argument being presented: "self driving is safer" is defeated by very obvious failures (driving into pillars, police cars, and people) that are /exactly/ the kinds of things that a computer should be able to handle. Human caused accidents in these situations are almost invariably due to inattentiveness that should not apply to any automated system.

So we get to the problem of "if it can't do handle the specific obvious case where it should be absolutely superior" (I would argue an automated system should not have ever been able to make either the police cruiser or the 280 crash in any circumstances. Those were both in perfect conditions - clear weather, etc, etc), how can we trust it to handle any of the complex situations that a self driving car necessarily must handle?

The Tesla system solves this by simply saying "the driver is still driving, we're merely assisting", but that assistance leads to inattentiveness - that's human nature, you have to design around it - which should be fine, if the systems were able to handle that, but that clearly are not.

I would argue that given the configuration of the Tesla system they are less safe, not necessarily due to net accidents, but rather because of the failure modes of the accidents. Bumper to bumper accidents on a freeway happen all the time, even driving perfectly you can end up in them because not everyone else is driving perfectly. The problem with things like the Tesla system is that it's generally fine, except in those cases where it drives at speed into objects because it gets confused.

If your vision or mental state leaves you unable to safely account for large obstructions you aren't allowed to drive. Because it isn't safe.